> For Codex, working with Jira and Atlassian was a hassle in my environment...
I hated this so much. Both acli, which my agents have to rediscover how to use from --help every time, and the MCP, which I have to reauthenticate against frequently. I have replaced both with a "skill" that just describes where to find an Atlassian API key and which version of the API to use. Works perfectly every time.
while everyone is somehow still stuck on and fascinated by claude, heres your quick update on the sota of coding models and harnesses mid august 2026
codex is good, both cli and desktop app, you get lots of usage on any plan.
sol is good! and gets the job done, write or dictate a very long and thoughtful prompt, and leave sol xhigh or max fast working on it for an hour or so
omp is an amazing harness, any feature claude code or codex is adding has likely already been here for a couple months.
good harness which im suggesting to all my developer friends, but for everyone else codex is the better option due to its simplicity and being the plug and play option
claude is decent, but not great. all models are somehow getting restrictive. you get basically unlimited opus on max plans, fable is good but slow and the random guardrails suck soo much which is why i havent used it once in weeks now.
gemini 3.7 is great for speed. everyone is sleeping on it, including even me
kimi k3 - great for frontend, one of the few models thats willing to commit crimes for you AND has the intelligence to have a chance at actually succeeding;
ds pro and flash are fast but not something id actually use for important things, unlike sol, fable and maybe 3.7 here and there
glm 5.3 i haven't tested yet
honorable mention to local models which are actually getting good now! 5090s will continue to get more and more expensive in the coming months. sadly.
theres way way more than claude in this world and its taking people surprisingly long to figure that out. maybe its for the best!
Pi by itself is more than capable, OMP is okay but you really don't need much for a great harness (these models are RL trained to hell to be a coding agent, sometimes less is more)
Fable/Sol/GLM 5.3/Kimi are its league (in that order)
Deepseek/Opus is solid
Qwen 27B is the floor - there's no reason to use Sonnet/Terra/Haiku
For everyday activity - I don't think you need to be using Sol (xhigh) for everything - unless you're made of money - I've found using Luna from OpenAI to be more than enough - it'll outreach to Opus/Sol when it needs to
Haven't had access to Gemini 3.7 but we're getting it at work soon, will give it a go!
Codex CLI is pretty bare bones in a bad way (at least Pi is extensible). Claude code is vibeslopped to the extreme
My token usage on Claude models has dropped by 83% over the last month - I'm pretty much only using it for quick one off questions or reading papers. it feels impossible for me to get Opus models to stop entering into cyclic loops, and my work is too security adjacent for Fable.
Codex has been an excellent workhorse - doesn't feel like I have to dance around the guardrails, doesn't lose _everything_ when it compacts, and doesn't litter the workspace with a million and one planning to plan files.
I have to agree with you there. I did some good work with Claude then Fable came out - impressed with that as well. Then they dropped access to it and upon returning was never the same - even the Opus models for some reason. Then one day I burned through my limit in about 10 minutes and had to get a project completed. I subscribed to Codex and it has been fantastic - finished my project and continued on to others. I just dropped my Claude max plan down to the pro and subscribed to the $200 plan on Codex.
I don’t get Claude, and that’s almost exactly what I did - I dropped to Claude Pro $20 + Codex Pro $100, and then unsubscribed from Claude and ramped up Codex. The Claude Pro is consumed within an hour on a simple task. I wish only Codex worked a bit faster than on the Fast mode.
I used to rely on Fable for research when it was first out, today it doesn’t seem to be much better than Opus, and it uses up the quota exceptionally fast - 1h Fable in a single short session, and there’s little left for Opus to hit the 5h limit in a second session. With Opus I get about 3-5h of relaxed use with a couple subagents to save the context, but there’s usually quite some disagreement between the subagents and orchestrator - Claude does some model routing with default agents and picks Haiku and Sonnet for subtasks - only later to disagree with them and redo the work - and burn extra tokens. With Claude, it’s really either Opus or Fable if you want some quality.
That said, their marketing is exceptionally effective. Virtually all nontech folks consider only Claude.
I don’t get how Claude is considered providing “unlimited” quotas. I use up my 5h on Max $100 and Team Premium in 2-3h of relaxed use of Opus 5 high+ on fresh sessions with just a couple skills/plugins. And my weekly quotas are gone in 3 days of such relaxed use. With Codex my $100 weekly quota is used up within 3 days with Sol high+ too.
Im not convinced to pay $200 for Claude’s models.
With Claude, I have to intervene every 15-20 minutes, it’s non-autonomous and it’s incredibly unreliable at self-correction. GPT is strong at self-correction but it tends to drift away from the plan to self-correct in a loop very often - a lot of tokens and time burnt on aimless churn. Opus tends to push its uninformed opinions and fake retrieval, drifting every turn increasingly farther from the intended and approved design. Opus skims over specs and makes too many mistakes.
As for closed frontier models, I prefer the GPT models over Claude’s.
I’ve started relying more on Grok, GLM, Kimi and DeepSeek models for subagents - I’ve ended up with a factory and am seeking to reduce my reliance on the closed frontier models - they’re just not SoTA on their own for development anymore.
Unfortunately the codex plans don’t offer the same amount of tokens as they did before. This changed around a week ago. There’s been a lot of user reports noticing this issue, and I’ve noticed the same pattern on my account. Previously I would never reach my weekly quota but last week I managed to finish it it one day. Same project, same single session sequential work. Not sure if there’s an issue or if it’s on purpose, and not even sure it applied to all accounts. Curious if other users on HN noticed the same problem.
It may be ipython making it work well with ds flash, too. I haven't really run many separate experiments, to be honest.
I also like prime-agent's way of handling sessions better than any other harness i've used. You can run multiple agents from one instance, although the scoping could be better.
But they can interact with past sessions, so preserving context isn't as important all the time. I just tell them to search for [thing] in another session.
It seems to have no problem with all the skills and things the other harnesses are using. I use superpowers and ponytail a lot.
It's my daily driver now. I like it better than opencode. But it doesn't ask permission. So I put it in a VM.
As a coda to this, anyone using grok 4.6 via API pricing should be aware that while their headline pricing is good, the pricing that actually matters is pretty bad.
Their cache read costs are $0.50 per million, or 25% of the cost of uncached reads.
The industry standard is a 90% discount, so cache costs you 10% of uncached. So that means 5.6 Sol actually costs less per million cache reads - $0.40/million.
If you are doing a lot of agentic work where the vast bulk of your token consumption will be cached input reads, you won't get the expected cost savings from Grok.
I imagine this is the result of some problem in their serving infrastructure that I hope they will fix, because then the pricing will become actually strong. (The other possibility is that they bet on distracting people with good headline prices assuming they'd miss the bad cache pricing, but I'll give them the benefit of the doubt on that.)
Speaking of crimes, the guy behind Grok not only manipulates it for propaganda purposes, but also he's the guy responsible for this: https://www.doge-impact.org/
Anyone using Sam Altman's OpenAI is making a poor ethical decision, but anyone using Grok is, objectively, supporting a monster.
By this logic everyone should have their own impact website. The suggestion that everyone right now not giving a meaningful percentage of their income to save a life is responsible for ending that life, is ridiculous.
Not everyone should have an impact website because not everyone is capable of causing 88 deaths per hour. Scale matters.
Take Flock for example. Reading license plate is legal. But when at done at scale, it's a massive loophole into violation of 4th amendment.
Based on how much energy average Americans use, maybe they are responsible for causing adverse effects elsewhere in the world. USAID could exist as a means to undo some of that. It does not anymore.
Calling another person a monster because you disagree with them (or what you heard about them from third parties) is not the pinnacle of civility. Just think about what you’re saying here. Monster: “Malformed animal or human, creature afflicted with a birth defect”. You don’t mean this literally, do you? You may want to spend a moment to think about what kind of company you’re putting yourself in with such wording and such thinking.
USAID was clearly contorted into a money laundering operation used to sway all kinds of things in bizarre directions as these things go. Arguing about whether there was some good or not is like complaining that SBF did some "Effective Altruism" during his crime spree. Shutting that down is one of the few sane things these jokers have done.
Let me give you an example of the money laundering operation. Due to USAID shutdown, Bangladesh went from ~$500M in US assistance to ~$71M, with bilateral health funding dropping ~97% in some analyses. Over 100 projects (~$550M) suspended overnight. 20k–50k development workers laid off (1,000+ at icddr,b, an award winning health research institution alone). TB programs (major USAID focus) largely halted. Bangladesh is high-burden; prior gains in case detection and falling death rates are at risk of reversing, plus higher chance of drug resistance from incomplete treatment. Immunization, maternal/child health, community clinics, nutrition, water/sanitation, and gender-based violence services sharply reduced. Child protection funding down ~36%. Food rations in Rohinhya camp, the largest refugee camp in the world, halved for >1M people; health and education services cut.
Now you can argue that US does not have any kind of obligation to send 500M to Bangladesh. But it sent it anyway, for years, and then DJT came and broke promises.
The inflated price you pay at gas station, groceries, and in interest when you're borrowing money, is a result of those broken promises.
Expecting an onslaught of cash as some permanent way of being, especially given the fickleness (and fragility) of any state let alone political regime is an incredibly daft move. I don't care if it's Europe or Israel or Bangladesh, all this is ultimately graft that comes back to bite the people taxed and sent to wars to enable it. You make an adjacent comment that insinuates the US economy is basically bunk, which means the free lunch is over anyway.
So it's a problem when a poverty striken nation expect aid to combat child mortality, but shelling out $150m on Juicero or $500m on Theranos is fine? Please try to answer without sounding like a psychopath.
Whatever point you are trying to make is not coming across, what even are these numbers and what do they have to do with citizenry of the United States? You also have a quantum view of the United States that it is and isn't impoverished, so it's supposed to liquidate to fund some other foreign entity that is not rate paying? I'm dizzy.
"Looking up synonyms" huh? You have yet to make a coherent argument and keep trying to attack the messenger, typical behavior when people get called out for false entitlement. Your "psychopathy" accusations are pure cowardice.
I’m sorry, but if I’m giving someone who is - at best - an acquaintance of mine $50 a month out of the goodness of my heart and then one day decide to stop, that’s not a broken promise. If that acquaintance got angry at me about stopping I’d get pretty upset back.
I really don’t understand what link you think there is between USAID spending being cut and inflation. Gas prices are obviously Iran. Everything else started years ago.
Inflation is high because interest rates are high. Interest rates are high because top holders of US Treasury bonds like Japan, UK, China, are all dumping bonds. Why do you think they're doing that?
> ds pro and flash are fast but not something id actually use for important things, unlike sol, fable and maybe 3.7 here and there
As someone who's used Gemini 3.7 Flash (Google sub mostly for the storage) and DS4 Flash a lot (~6B tokens), I'd actually place DS4 Flash (even pre-0713) above Gemini 3.7 Flash. Gemini has a tendency to leave some things unimplemented; perhaps it's agy which frankly leaves a bit to be desired as a harness.
Although I will praise DS4 Flash any day, it no longer makes sense for me after the price increase (GPT 5.6 Luna is a much better price point) and I have completely migrated my high volume workflows to Muse Spark 1.2 Contributor (which I find to perform better than DS4 Flash 0713, happily).
the future is here and one should be thankful for its slightly uneven distribution. otherwise we would hardly have anything left about which to develop strong opinions!
> gemini 3.7 is great for speed. everyone is sleeping on it
Is this Gemini 3.7 Flash by any chance? Then - No. Not sleeping on it. It’s just not good.
I had a Python package build fail this week due to an unpinned dependency. Gave it to Gemini spent 5-7mins before I noticed it going off in some tangent. Reran with Claude Opus 4.8 - fixed in under a minute.
I know anecdata of one. But something like this has happened every time I test a new model from Google.
I only have the $20 claude subscription. Last weekend I was doing a fairly heavy task (porting quake to raspberry pi native GLES 1.0) and ran out of claude usage twice, then hooked in with Luna in OpenCode and finished everything off for like $0.40 in tokens, which was impressive.
With Opus 5.0 being kinda crappy vs 4.8, I think Anthropic is in trouble.
On the flip side, I've been using Fable to statically recompile a game binary from a system that's never been emulated to modern C++ while also keeping the code clean, maintainable, and portable and it's been a dream. I'm in awe of how fast Fable is able to bring a 20 year old game that's largely been lost to the sands of time to SDL3. It finished the recomp in about 3 hours. The next day has been fully rewriting the functions to remove old hardware-specific quirks and then rewriting large swaths into multiple classes, subsystems, building tests, etc. Tomorrow I'm going to add mod support and then experiment with a random platform like making a Switch build of the game.
It's expensive but it's doing in hours what no one's done in 2 decades.
iPod Classic click wheel games. Specifically Mini Golf. I've also been working on an emulator for the others and it can now boot all of the 20 decrypted games released ~20 years ago and play most of them. There's a Sims Bowling, Sims Pool, and Lost game in there which is cool from a preservation standpoint. Less cool are the SAT Prep 2008 games which have actually been surprisingly annoying to emulate due to the way the text uses blend modes.
I plan on releasing all of this at one point. It's crazy it hasn't been done in 20 years!
Model performance is very much subjective to what you’re using it for. I’m a PM and doing mostly knowledge work, and I’ve been really happy with Opus 5 as my daily driver with very good results, perhaps my favorite anthropic model so far. I do a lot of front-end coding with it too. I prefer it over the openAI models, been switching frequently. Kimi K3 is great too.
This post needs an edit. Author is not comparing "Codex" and "Claude". They are comparing Codex TUI/CLI with (presumably) gpt-5.6-sol, against Claude Code TUI/CLI with (presumably) Claude-Opus-5.
Ctrl + f > [5.6, sol, sonnet, opus or fable] yields no results.
"Claude" is a product family, which includes Models, and Harnesses (and probably more). "Claude code" covers both the Claude Code TUI, and CC in the Claude desktop app.
"Codex" is the same, and could refer to the Codex TUI, or Codex in the ChatGPT (formerly codex) desktop app. (And well, historically, gpt-5.*-codex.)
Hearing "Yea Claude is great for coding" takes an hour off my life.
Something something "Honey why don't you finish up with your Nintendo and come to dinner?"
I’m sure it depends on the type of work, but for mine, Codex is much more helpful. Honestly, it mostly comes down to it being significantly faster, probably because as many have said, it seems tuned to not spit out word vomit, both in its chat interactions, and its code (Claude is obsessed with massive comment blocks that are basically guaranteed to become dead context noise if you ever use it to iterate on code).
I mostly do very obsessive, tightly scoped, carefully thought out small changes on a fairly boring stack, one interaction at a time, verifying functionality and code. I know what I am doing, but I also know what I don’t like doing (the same exact set of things I’ve already done a dozen times in my career)
> Codex feels more like a version of Data from Star Trek
Great analogy for some reason. At fist I felt Codex Sol was a bit more cold. But now that I've worked with it for several weeks it has grown on me, even shown some personality. I appreciate that it is a bit more business-like, Fable is a bit too friendly sometimes when it ought to be focused on work. Codex can be a bit more nit-picky.
I agree with most of his other observations. I've already started to bin tasks based on which model I feel is best suited. In general, for well scoped and straight ahead tasks where banging out code is what I want I reach for Codex. For less specced tasks where I need a broader view and want the model to fill in more details I reach for Fable.
Both are great and they make a good team together.
I cancelled my Claude subscription to give Codex a proper shot for a month. 1 week in, I'm more than happy. Sol is fast and capable. Luna as subagents for execution are very cheap usage wise (and API price wise too).
I think more than anything else, I don't get a headache conversing with Sol. That alone is enough reason for me to stick to Codex.
Experimenting with adding open source models to the mix to get more execution done while using Sol as the brain.
> It felt to me that Codex created a much simpler solution in terms of code architecture than Claude.
Wow, I made exactly the opposite experience. Codex loves to make things as complicated as possible, even ignoring instructions and predefined skills. Claude behaves way more pragmatic. Maybe depends on the type of work one does, or even which programming languages/frameworks are used?
I have the same experience. I have passed Codex's code over to Claude (Sonnet mostly) to simplify it and it did a so much better job. Codex couldn't comprehend a simpler solution no matter what.
Overall I prefer Codex over Claude in their current state, but I do agree that Codex's solutions tend to be too complex and I have to constantly tell it to simplify solutions, detect when it's creating ornate architecture, and proposing data models that have way too much going on. In my usage that's its main weakness.
Sol is for routine work, Opus for frontend/design, and Fable for more complex / ambiguous / architecture work. Fable works extremely well to drive Sol as a subagent.
Fable is the only one you can actually trust to not look at the code, but Sol is somehow still more pleasant to work with, especially in fast mode. Opus is the enemy, and it will make you insane if you talk to it for too long.
When discussing Claude vs. Codex, etc. I find it necessary to make the distinction between the models and the harness.
Claude's models in my experience do a better job of inferring my intent, or to say it does a better job of giving me the result I imagined in my mind. A recent example was a UI prototype I was building for a desktop application. I had asked GPT's 5.6 Sol to update the open document in the prototype to better reflect the context of the feature I was designing, and 5.6 Sol took it very literally and had just added some text to the currently open document, not what I had in mind. I tried again with Claude Opus 5 and it added a completely new tab with a complete new document that, although imperfect, much better matched my expectations.
You could say this was a prompting skill issue, but seeing how many people are prompting their AI I believe the labs are incentivized to continue to improve their ability to infer intent.
When it comes to the desktop applications though, I find Claude Desktop's output to be incredibly verbose and full of jargon. I feel like it hits me with an entire essay and the UI doesn't have enough typographic hierarchy to make it easy to scan. ChatGPT Desktop is much better in this regard, I feel the output is concise, clear, and gives me just enough info to feel in the loop without being overwhelmed. Even though I have the setting on for technical language, it feels more understandable than Claude. I also feel that ChatGPT's desktop app has a better design and much more polish.
I do not really like how bloated both applications have become though. This weird segmentation of Chat, Work, and Code all just seems like it's pushing a technical limitation onto the user. The other day I opened a document in ChatGPT and asked it to do something, then it told me it could only do it in work "mode", so it then created an entirely new conversation with a reference to the previous conversation. It wasn't a completely new area of the UI either, it just added a "Work" badge to the new conversation in the list. Feels a bit unnecessary, like couldn't you just keep it all within the same conversation?
For me, codex $100 mo/plan and a claude teams account at work (mostly sonnet, some opus), Claude basically feels about as effective as Codex did 4-5 months ago pre-5.6. Claude still has weird patterns of being confident in one answer while another chat with the same model is confident in another answer, where one answer is clearly wrong. Missed details, over-engineering in places, while still overall helpful and effective. Codex, however, just feels freaking rock solid on Sol high. I literally have zero complaints.
guys, everything is about the harness, not the model! I even use Hy3 for free from nous portal, yes, it finishes the task in 50 minutes instead of 5, as Terra of codex would do, but I have it under the same harness (Hermes Agent) so result is more or less the same and it costed me nothing and worked at background while I was doing other things :)
I believe that “it’s about the harness” captures an important idea, but I also think it’s helpful to consider the harness and the procedure separately.
The harness provides the model with its tools, context, and environment for execution.
However, I prefer not to have the model within that harness also bear the responsibility for remembering the process steps—like plan, implement, review, fix, and verify—deciding when to move from one to the next, and keeping track of the loop's state.
For tasks that need to be repeated, I’ve been moving that part into a reliable, deterministic runtime. Inside it, Claude, Codex etc simply take on interchangeable roles.
This setup makes changing models much simpler: the overall process remains consistent, and each role can be optimized independently.
"How this article was written
I wrote this article and used Grammarly to proofread and fix it."
What a brave new world we're in, where this is necessary. Regardless, it's appreciated. Although, I have the feeling that those using an LLM to do most of their writing will be less likely to include such a disclaimer.
I’ve been experimenting with this for a while, and right now I’m using Luna xhigh as my default. Previously I was using Sol medium.
Sol medium is a great balance between speed and being thorough, but it’s quite expensive. Luna xhigh seems to compensate for slightly lower intelligence by thinking and reasoning for longer, so tasks can take more time to complete. But it’s crazy cheap.
I also have some custom evals using promptfoo to make sure I’m not introducing regressions when switching models. So far, Luna xhigh has been really, really good for the price.
Claude can add comments in code that are better just living as a comment on the PR.
You'll ask it to do something and it'll comment the code with an answer to what you asked it, rather than just explanatory comments to whoever comes after.
There's also a second issue that if the code is actually incorrect, the comment can nevertheless bolster the case for it.
I found that Claude would litter the codebase with what felt like notes-to-self. Sometimes it would make an abstraction that wasn't very well thought out and when I pushed back on this choice, it would usually go with something more along the lines of what I suggested in my pushback, but left a comment like "use the git CLI instead of re-implementing git" or something.
It felt like it was commenting on the diff sometimes instead of what the code was doing.
Those verbose comments provide context for why something is there way it is, so it can take those decisions in account when making changes in the future.
It writes out stories describing what isn't there or what used to be there. It's usually not helpful, just noise. It also likes to write it in very verbose AI-styled prose.
And the way they are written is rather unhelpful: '//this not cosmetic: this css helps the user understand where he should click'. The 'this is not' pattern is useless at best.
It's likely not helpful to you, but it is to Claude. Basically a memory entry at the exact point of relevance, with the same caveats. One advantage though is it survives cloning, so another using the project benefits.
Confusion usually surfaces immediately, so it can be steered properly. Being sure of a decision that turns out to be wrong usually means a bug that could bite when you least expect it.
It does visibly use a few more tokens in the short term. But it hurts Claude's ability in the long term if removed because it has to either think more every time it encounters that section in the code (more invisible thinking tokens used), or there's an increased risk of making a maladaptive change without context. Heck I've been bitten enough times already by it, though more recently it's cross project concern where Claude rediscovers some root cause that was already found and documented (I have it save a centrally-located writeup for things that took serious work to learn, when I remember, but I still need to make that knowledge pool known to other agents/sessions without having to manually prompt each time) elsewhere.
Problem is today's LLMs don't have the long term memory that humans have, and so remembering the reason behind a given change/decision has to be preserved in some way if it's non-obvious. Hence why there is {AGENTS|CLAUDE}.md, the auto-memory system, and 1001 variants of memory implementations in the wild. All are trying to ensure that LLMs can have the context they need at the location and time they need it. And you want to block Claude from using a technique that it natively finds helpful.
If they have in mind what I think of when I think of claude comments then frequently they are only relevant to the back and forth that produced the code and are exaggerated to the point of being misleading. A very common example is to land on something that isn't quite right or doesn't quite work, prompt an adjustment, and then the adjustment will have some grandiose prose about "preventing the critical bug that shipped once." People just need to clean up such litter manually but it would be nice if the models generated less of it.
The story of how the code got to a certain state belongs primarily in git commit messages, not in comments. Comments should serve to aid in understanding of the code, highlight footguns or hidden dependencies, etc.
Claude very often litters code with comments about decisions that were made within a single session/pull request, its just noise.
If one needs to dig history, then git log -p and similar are way more useful and reflect the real history rather than LLM inference about it. Comments reflecting the goal are more useful as neither history nor code itself covers it.
Looking into git history is an active decision that's usually only done after something undesirable happens, and now there's the need to try and find why. Better to have the context passively loaded as part of working with the relevant section of the code, and likely prevent that something from even happening in the first place. Think Chesterton Fence.
Those massive comments generally refer to intermediate states that never shipped, making them utterly useless, noisy and confusing as a permanent artifact. They're sometimes useful to reviewers, so a commit message or pull request comment is much more appropriate.
That's your perspective. For Claude that's an extension of its thinking, which makes it work better. Just like the person who takes notes so they have references for later. Take it away and you're negatively impacting outcomes.
No, it's the other way around. If you put a comment "don't do foo" in your code, you're biasing future readers, human or artificial, to think about foo when you explicitly want them not to.
It isn't about making readers not think about foo. It's about biasing thought about foo in a particular direction. And I've seen it work multiple times: agent mentions it encountered X comments in the code and revised its action as a result (yes, I read live agent transcripts).
I've been using Zed mostly with Codex session windows and its great. It is much easier to handle multiple sessions than in the Codex VSCode app and the interface is quite good.
I find Claude more often gets my intent without having to spell things out for it, while Codex gets hung up on minor details and over-engineers a solution for them.
I think a part of this is that people tend to undervalue their own skills and expertise when talking about these anecdotes.
A lot of people in the comments do have a software engineering background. People at different skill levels in different backgrounds are going to be using these tools in different ways, and that's going to heavily impact their experiences with these models.
Sure, there are differences between Fable and Sol. But I've even seen people on here saying that they're getting better mileage out of Qwen models they're self hosting.
I think the driver is just as important than the car, when it comes to this sort of stuff.
I think this definitely applies to fable/opus/sol, mixed with the undeterministic nature of the models. But those claiming Qwen are just outright coping. Thats nonsense imo
I've noticed that too with Sol (xhigh) it starts with pretty good architectural goals, but really get lost in the weeds with certification and validation. Like it won't trust native tools like (in my case packer, Ansible, gcloud) to fail reliably under possible error conditions, and insists on implementing custom verification.
It also doesn't have a clear idea of what the actual threat model is, and builds all kinds of extremely defensive systems to account for imagined hostile actors. I'm like "Dude, it's only our systems that are creating these SVGs, they're never going to be user supplied, so you don't need to write an entire validation and sanitation framework here."
It also seems to treat the desired initial state of something as a permanent invariant and designs elaborate tests to ensure that it remains that way. Then when you make one little change it has to go and update a ton of tests it created.
I've had to rip out a bunch of overengineered jank from several feature implementations, and in doing so I ended up having to create retrospective documents that warn against this kind of behavior that I'll have the model review whenever a plan begins to go sideways.
What, you don't like it proposing sha256 checksums and state enums for everything?? :D
It does seem to have this predilection, but I find it calms down if I tell it to simplify for "this phase". Sure, we'll pick that up down the road.. I promise. :D
The speed is the first big contrast; I have a routine multi-step skill that I run several of per week. Opus 5 was routinely taking 2 hours to do it, while older Claude models took around 20 mins; Codex restored that speed.
Second is legibility. Somebody wrote in one of the related discussions yesterday that Claude's current linguistic contortions could legitimately be considered damaging to mental health, which doesn't seem (too) hyperbolic to me. Codex (Sol) isn't perfect but it's much more direct. And so far I haven't seen it display much of an attitude, vs Opus's infuriating passive aggressive sulky know it all personality.
I slightly prefer Anthropic to OpenAI as a company, but I will vote with my wallet and discontinue my max subscription unless Anthropic does some serious damage control within the next week or two.
I broadly agreed with the authors experience, although I wouldn't say codex does anything "wrong". I think different agents/LLM's have different personalities, and it takes learning to understand how to get them to do what you want. I had the same experience when we started using claude at work, and I was trained on codex. Claude seemed to do everything "wrong", because I was writing instructions designed for codex.
I do agree claude looks for more things to do in your repo, whereas codex is more likely to do what its old and stop. Which is better is personal preference as far as I can tell.
I honestly would've used Claude more if not for the confusing presentation/font, I'm not an expert on this subject, Codex is just a lot easier on my eyes.
>> I think the main difference I feel between Claude and Codex is that Claude tries to go above and beyond what is asked and guess what you might want and then directly do it, while Codex is more like a companion that does what you tell it but will not overdo it. It will stop at the first sign that it might be done.
tl;dr I gave GPT 5.6 a small-medium sized ticket, which should have been several hundred lines plus tests. It ended up creating a 25,000+ line diff. Another GPT 5.6 Sol with fresh context looked at the worktree and said 98% of it should be thrown away. Claude thought the same, and suggested that several dozen compactions the model went through over several hours must have caused it to go adrift. I guess that's one consequence of having a relatively small context window.
I still use Sol quite a bit. I find that it's consistently the opposite of what the author describes: it's too relentless. It doesn't know when to stop. Opus is the opposite: it'll give up a bit too easily. If everything goes well that's not an issue, but often times it'll say things like "task is done, btw I couldn't do X Y Z" and X Y Z will be some important verification step that failed because another agent was using that resource or something.
At this point I trust GPT 5.6 mostly with surgical changes, or general codebase exploration tasks. It is a faster model, so it's easier to get small things done with it. For everything else I prefer Claude, despite its annoying tendencies.
you hear things like this about every model. they're probabilistic and you can't trust them, although i doubt that kind of odd behavior would happen on low or medium effort levels. i find all of the models today pretty good at following instructions generally speaking. i'm sure you did /plan, but i rarely see it stray very far from the plan. Then again, I never walk away, since i'm always spinning up another task that can be done in parallel, and I don't let things run overnight because I'm not made of token cash (yet)
Claude Code seems more generous with its quota, which is why I use it as my main driver.
That said, Codex does seem more capable, terse, and faster. There are some tasks that Claude can't handle but Codex can. One example was a WinForms binding/project deserialization bug. Sorry, the code is a mess, so even I couldn't quite figure out which part was causing which problem.
I initially thought the bug would be difficult to reproduce in a unit-test setting. Claude could only narrow down the problem and tell me where to put a breakpoint. Codex, on the other hand, actually managed to create a reproducible unit test first, and then used that to fix the bug. That impressed me.
The only problem is the quota. Codex burns through it very, very quickly, even when I'm just using Terra 5.6 Medium. That's basically why Claude Code remains my main driver despite Codex seeming more capable.
That's surprising to me. I recently switched from Claude to Codex because I was blowing through my Claude limits. With Codex I haven't been able to use up all of my usage on the 20x plan, with Claude and fable I could do that in a day. Codex is also a lot faster than Claude. I do think Fable is slightly better and can come up with better abstractions than Codex but trying to read it's output became really frustrating to me.
> For Codex, working with Jira and Atlassian was a hassle in my environment...
I hated this so much. Both acli, which my agents have to rediscover how to use from --help every time, and the MCP, which I have to reauthenticate against frequently. I have replaced both with a "skill" that just describes where to find an Atlassian API key and which version of the API to use. Works perfectly every time.
while everyone is somehow still stuck on and fascinated by claude, heres your quick update on the sota of coding models and harnesses mid august 2026
codex is good, both cli and desktop app, you get lots of usage on any plan. sol is good! and gets the job done, write or dictate a very long and thoughtful prompt, and leave sol xhigh or max fast working on it for an hour or so
omp is an amazing harness, any feature claude code or codex is adding has likely already been here for a couple months. good harness which im suggesting to all my developer friends, but for everyone else codex is the better option due to its simplicity and being the plug and play option
claude is decent, but not great. all models are somehow getting restrictive. you get basically unlimited opus on max plans, fable is good but slow and the random guardrails suck soo much which is why i havent used it once in weeks now.
gemini 3.7 is great for speed. everyone is sleeping on it, including even me
kimi k3 - great for frontend, one of the few models thats willing to commit crimes for you AND has the intelligence to have a chance at actually succeeding;
ds pro and flash are fast but not something id actually use for important things, unlike sol, fable and maybe 3.7 here and there
glm 5.3 i haven't tested yet
honorable mention to local models which are actually getting good now! 5090s will continue to get more and more expensive in the coming months. sadly.
theres way way more than claude in this world and its taking people surprisingly long to figure that out. maybe its for the best!
Pi by itself is more than capable, OMP is okay but you really don't need much for a great harness (these models are RL trained to hell to be a coding agent, sometimes less is more)
I run a lot of SlopCodeBench - https://github.com/michaelasper/benchmarks
Fable/Sol/GLM 5.3/Kimi are its league (in that order) Deepseek/Opus is solid Qwen 27B is the floor - there's no reason to use Sonnet/Terra/Haiku
For everyday activity - I don't think you need to be using Sol (xhigh) for everything - unless you're made of money - I've found using Luna from OpenAI to be more than enough - it'll outreach to Opus/Sol when it needs to
Haven't had access to Gemini 3.7 but we're getting it at work soon, will give it a go!
Codex CLI is pretty bare bones in a bad way (at least Pi is extensible). Claude code is vibeslopped to the extreme
My token usage on Claude models has dropped by 83% over the last month - I'm pretty much only using it for quick one off questions or reading papers. it feels impossible for me to get Opus models to stop entering into cyclic loops, and my work is too security adjacent for Fable.
Codex has been an excellent workhorse - doesn't feel like I have to dance around the guardrails, doesn't lose _everything_ when it compacts, and doesn't litter the workspace with a million and one planning to plan files.
I have to agree with you there. I did some good work with Claude then Fable came out - impressed with that as well. Then they dropped access to it and upon returning was never the same - even the Opus models for some reason. Then one day I burned through my limit in about 10 minutes and had to get a project completed. I subscribed to Codex and it has been fantastic - finished my project and continued on to others. I just dropped my Claude max plan down to the pro and subscribed to the $200 plan on Codex.
I don’t get Claude, and that’s almost exactly what I did - I dropped to Claude Pro $20 + Codex Pro $100, and then unsubscribed from Claude and ramped up Codex. The Claude Pro is consumed within an hour on a simple task. I wish only Codex worked a bit faster than on the Fast mode.
I used to rely on Fable for research when it was first out, today it doesn’t seem to be much better than Opus, and it uses up the quota exceptionally fast - 1h Fable in a single short session, and there’s little left for Opus to hit the 5h limit in a second session. With Opus I get about 3-5h of relaxed use with a couple subagents to save the context, but there’s usually quite some disagreement between the subagents and orchestrator - Claude does some model routing with default agents and picks Haiku and Sonnet for subtasks - only later to disagree with them and redo the work - and burn extra tokens. With Claude, it’s really either Opus or Fable if you want some quality.
That said, their marketing is exceptionally effective. Virtually all nontech folks consider only Claude.
I don’t get how Claude is considered providing “unlimited” quotas. I use up my 5h on Max $100 and Team Premium in 2-3h of relaxed use of Opus 5 high+ on fresh sessions with just a couple skills/plugins. And my weekly quotas are gone in 3 days of such relaxed use. With Codex my $100 weekly quota is used up within 3 days with Sol high+ too.
Im not convinced to pay $200 for Claude’s models.
With Claude, I have to intervene every 15-20 minutes, it’s non-autonomous and it’s incredibly unreliable at self-correction. GPT is strong at self-correction but it tends to drift away from the plan to self-correct in a loop very often - a lot of tokens and time burnt on aimless churn. Opus tends to push its uninformed opinions and fake retrieval, drifting every turn increasingly farther from the intended and approved design. Opus skims over specs and makes too many mistakes.
As for closed frontier models, I prefer the GPT models over Claude’s.
I’ve started relying more on Grok, GLM, Kimi and DeepSeek models for subagents - I’ve ended up with a factory and am seeking to reduce my reliance on the closed frontier models - they’re just not SoTA on their own for development anymore.
Unfortunately the codex plans don’t offer the same amount of tokens as they did before. This changed around a week ago. There’s been a lot of user reports noticing this issue, and I’ve noticed the same pattern on my account. Previously I would never reach my weekly quota but last week I managed to finish it it one day. Same project, same single session sequential work. Not sure if there’s an issue or if it’s on purpose, and not even sure it applied to all accounts. Curious if other users on HN noticed the same problem.
I think harness/model pairs matter more than your analysis lets on.
I've had great luck with the ds flash v4, paired with prime-agent for the harness--I like the results a lot. And you get to see thinking tokens.
I haven't liked the model as much in opencode.
Sol & luna have been great everywhere. sol plans, luna builds.
Prime Agent looks really interesting. Both the "recursive language model" bit and routing everything through IPython.
https://github.com/PrimeIntellect-ai/prime-agent
It may be ipython making it work well with ds flash, too. I haven't really run many separate experiments, to be honest.
I also like prime-agent's way of handling sessions better than any other harness i've used. You can run multiple agents from one instance, although the scoping could be better.
But they can interact with past sessions, so preserving context isn't as important all the time. I just tell them to search for [thing] in another session.
It seems to have no problem with all the skills and things the other harnesses are using. I use superpowers and ponytail a lot.
It's my daily driver now. I like it better than opencode. But it doesn't ask permission. So I put it in a VM.
Not mentioning Grok 4.6 here is a crime. Fast and accurate.
And it can communicate, unlike the gobbledygook that comes out of Claude.
As a coda to this, anyone using grok 4.6 via API pricing should be aware that while their headline pricing is good, the pricing that actually matters is pretty bad.
Their cache read costs are $0.50 per million, or 25% of the cost of uncached reads.
The industry standard is a 90% discount, so cache costs you 10% of uncached. So that means 5.6 Sol actually costs less per million cache reads - $0.40/million.
If you are doing a lot of agentic work where the vast bulk of your token consumption will be cached input reads, you won't get the expected cost savings from Grok.
I imagine this is the result of some problem in their serving infrastructure that I hope they will fix, because then the pricing will become actually strong. (The other possibility is that they bet on distracting people with good headline prices assuming they'd miss the bad cache pricing, but I'll give them the benefit of the doubt on that.)
> Not mentioning Grok 4.6 here is a crime.
Not yet. Don't give the guy ideas.
Been working a lot recently with Grok 4.6 for implementation and gpt 5.6 sol for review. Worked really good so far.
Why not the other way around?
I would not use Grok if it paid me per token… wild wild stuff…
Speaking of crimes, the guy behind Grok not only manipulates it for propaganda purposes, but also he's the guy responsible for this: https://www.doge-impact.org/
Anyone using Sam Altman's OpenAI is making a poor ethical decision, but anyone using Grok is, objectively, supporting a monster.
I can't wait until you research the people responsible for Chinese models....
By this logic everyone should have their own impact website. The suggestion that everyone right now not giving a meaningful percentage of their income to save a life is responsible for ending that life, is ridiculous.
Not everyone should have an impact website because not everyone is capable of causing 88 deaths per hour. Scale matters.
Take Flock for example. Reading license plate is legal. But when at done at scale, it's a massive loophole into violation of 4th amendment.
Based on how much energy average Americans use, maybe they are responsible for causing adverse effects elsewhere in the world. USAID could exist as a means to undo some of that. It does not anymore.
Nobody cares bro.
Calling another person a monster because you disagree with them (or what you heard about them from third parties) is not the pinnacle of civility. Just think about what you’re saying here. Monster: “Malformed animal or human, creature afflicted with a birth defect”. You don’t mean this literally, do you? You may want to spend a moment to think about what kind of company you’re putting yourself in with such wording and such thinking.
The person you are replying to maybe should have better referred to him as having “no moral compass”, which I believe is quite accurate.
USAID was clearly contorted into a money laundering operation used to sway all kinds of things in bizarre directions as these things go. Arguing about whether there was some good or not is like complaining that SBF did some "Effective Altruism" during his crime spree. Shutting that down is one of the few sane things these jokers have done.
Let me give you an example of the money laundering operation. Due to USAID shutdown, Bangladesh went from ~$500M in US assistance to ~$71M, with bilateral health funding dropping ~97% in some analyses. Over 100 projects (~$550M) suspended overnight. 20k–50k development workers laid off (1,000+ at icddr,b, an award winning health research institution alone). TB programs (major USAID focus) largely halted. Bangladesh is high-burden; prior gains in case detection and falling death rates are at risk of reversing, plus higher chance of drug resistance from incomplete treatment. Immunization, maternal/child health, community clinics, nutrition, water/sanitation, and gender-based violence services sharply reduced. Child protection funding down ~36%. Food rations in Rohinhya camp, the largest refugee camp in the world, halved for >1M people; health and education services cut.
Now you can argue that US does not have any kind of obligation to send 500M to Bangladesh. But it sent it anyway, for years, and then DJT came and broke promises.
The inflated price you pay at gas station, groceries, and in interest when you're borrowing money, is a result of those broken promises.
Expecting an onslaught of cash as some permanent way of being, especially given the fickleness (and fragility) of any state let alone political regime is an incredibly daft move. I don't care if it's Europe or Israel or Bangladesh, all this is ultimately graft that comes back to bite the people taxed and sent to wars to enable it. You make an adjacent comment that insinuates the US economy is basically bunk, which means the free lunch is over anyway.
So it's a problem when a poverty striken nation expect aid to combat child mortality, but shelling out $150m on Juicero or $500m on Theranos is fine? Please try to answer without sounding like a psychopath.
Whatever point you are trying to make is not coming across, what even are these numbers and what do they have to do with citizenry of the United States? You also have a quantum view of the United States that it is and isn't impoverished, so it's supposed to liquidate to fund some other foreign entity that is not rate paying? I'm dizzy.
It's probably not coming across because you're looking up too many synonyms.
"Looking up synonyms" huh? You have yet to make a coherent argument and keep trying to attack the messenger, typical behavior when people get called out for false entitlement. Your "psychopathy" accusations are pure cowardice.
I’m sorry, but if I’m giving someone who is - at best - an acquaintance of mine $50 a month out of the goodness of my heart and then one day decide to stop, that’s not a broken promise. If that acquaintance got angry at me about stopping I’d get pretty upset back.
I really don’t understand what link you think there is between USAID spending being cut and inflation. Gas prices are obviously Iran. Everything else started years ago.
Inflation is high because interest rates are high. Interest rates are high because top holders of US Treasury bonds like Japan, UK, China, are all dumping bonds. Why do you think they're doing that?
> ds pro and flash are fast but not something id actually use for important things, unlike sol, fable and maybe 3.7 here and there
As someone who's used Gemini 3.7 Flash (Google sub mostly for the storage) and DS4 Flash a lot (~6B tokens), I'd actually place DS4 Flash (even pre-0713) above Gemini 3.7 Flash. Gemini has a tendency to leave some things unimplemented; perhaps it's agy which frankly leaves a bit to be desired as a harness.
Although I will praise DS4 Flash any day, it no longer makes sense for me after the price increase (GPT 5.6 Luna is a much better price point) and I have completely migrated my high volume workflows to Muse Spark 1.2 Contributor (which I find to perform better than DS4 Flash 0713, happily).
Claude code seems like a beginner's trap at this point.
the future is here and one should be thankful for its slightly uneven distribution. otherwise we would hardly have anything left about which to develop strong opinions!
> gemini 3.7 is great for speed. everyone is sleeping on it
Is this Gemini 3.7 Flash by any chance? Then - No. Not sleeping on it. It’s just not good.
I had a Python package build fail this week due to an unpinned dependency. Gave it to Gemini spent 5-7mins before I noticed it going off in some tangent. Reran with Claude Opus 4.8 - fixed in under a minute.
I know anecdata of one. But something like this has happened every time I test a new model from Google.
I only have the $20 claude subscription. Last weekend I was doing a fairly heavy task (porting quake to raspberry pi native GLES 1.0) and ran out of claude usage twice, then hooked in with Luna in OpenCode and finished everything off for like $0.40 in tokens, which was impressive.
With Opus 5.0 being kinda crappy vs 4.8, I think Anthropic is in trouble.
On the flip side, I've been using Fable to statically recompile a game binary from a system that's never been emulated to modern C++ while also keeping the code clean, maintainable, and portable and it's been a dream. I'm in awe of how fast Fable is able to bring a 20 year old game that's largely been lost to the sands of time to SDL3. It finished the recomp in about 3 hours. The next day has been fully rewriting the functions to remove old hardware-specific quirks and then rewriting large swaths into multiple classes, subsystems, building tests, etc. Tomorrow I'm going to add mod support and then experiment with a random platform like making a Switch build of the game.
It's expensive but it's doing in hours what no one's done in 2 decades.
Which one?
iPod Classic click wheel games. Specifically Mini Golf. I've also been working on an emulator for the others and it can now boot all of the 20 decrypted games released ~20 years ago and play most of them. There's a Sims Bowling, Sims Pool, and Lost game in there which is cool from a preservation standpoint. Less cool are the SAT Prep 2008 games which have actually been surprisingly annoying to emulate due to the way the text uses blend modes.
I plan on releasing all of this at one point. It's crazy it hasn't been done in 20 years!
Model performance is very much subjective to what you’re using it for. I’m a PM and doing mostly knowledge work, and I’ve been really happy with Opus 5 as my daily driver with very good results, perhaps my favorite anthropic model so far. I do a lot of front-end coding with it too. I prefer it over the openAI models, been switching frequently. Kimi K3 is great too.
Opus 5 is trash imo. 4.8 was better. 5 is overly verbose and complicates solutions. I switched to sol for my coding since and have been pretty happy
Im trying to do my part in the world here.
This post needs an edit. Author is not comparing "Codex" and "Claude". They are comparing Codex TUI/CLI with (presumably) gpt-5.6-sol, against Claude Code TUI/CLI with (presumably) Claude-Opus-5.
Ctrl + f > [5.6, sol, sonnet, opus or fable] yields no results.
"Claude" is a product family, which includes Models, and Harnesses (and probably more). "Claude code" covers both the Claude Code TUI, and CC in the Claude desktop app.
"Codex" is the same, and could refer to the Codex TUI, or Codex in the ChatGPT (formerly codex) desktop app. (And well, historically, gpt-5.*-codex.)
Hearing "Yea Claude is great for coding" takes an hour off my life.
Something something "Honey why don't you finish up with your Nintendo and come to dinner?"
i was going to respond to you with a snarky holier-than-thou remark but then read the article and holy cow, nothing is specified at all…
really feels like discussion spawns only off post title and as a second or third order effect, post content
I’m sure it depends on the type of work, but for mine, Codex is much more helpful. Honestly, it mostly comes down to it being significantly faster, probably because as many have said, it seems tuned to not spit out word vomit, both in its chat interactions, and its code (Claude is obsessed with massive comment blocks that are basically guaranteed to become dead context noise if you ever use it to iterate on code).
I mostly do very obsessive, tightly scoped, carefully thought out small changes on a fairly boring stack, one interaction at a time, verifying functionality and code. I know what I am doing, but I also know what I don’t like doing (the same exact set of things I’ve already done a dozen times in my career)
> Codex feels more like a version of Data from Star Trek
Great analogy for some reason. At fist I felt Codex Sol was a bit more cold. But now that I've worked with it for several weeks it has grown on me, even shown some personality. I appreciate that it is a bit more business-like, Fable is a bit too friendly sometimes when it ought to be focused on work. Codex can be a bit more nit-picky.
I agree with most of his other observations. I've already started to bin tasks based on which model I feel is best suited. In general, for well scoped and straight ahead tasks where banging out code is what I want I reach for Codex. For less specced tasks where I need a broader view and want the model to fill in more details I reach for Fable.
Both are great and they make a good team together.
I cancelled my Claude subscription to give Codex a proper shot for a month. 1 week in, I'm more than happy. Sol is fast and capable. Luna as subagents for execution are very cheap usage wise (and API price wise too).
I think more than anything else, I don't get a headache conversing with Sol. That alone is enough reason for me to stick to Codex.
Experimenting with adding open source models to the mix to get more execution done while using Sol as the brain.
> It felt to me that Codex created a much simpler solution in terms of code architecture than Claude.
Wow, I made exactly the opposite experience. Codex loves to make things as complicated as possible, even ignoring instructions and predefined skills. Claude behaves way more pragmatic. Maybe depends on the type of work one does, or even which programming languages/frameworks are used?
I have the same experience. I have passed Codex's code over to Claude (Sonnet mostly) to simplify it and it did a so much better job. Codex couldn't comprehend a simpler solution no matter what.
Overall I prefer Codex over Claude in their current state, but I do agree that Codex's solutions tend to be too complex and I have to constantly tell it to simplify solutions, detect when it's creating ornate architecture, and proposing data models that have way too much going on. In my usage that's its main weakness.
This is a good idea. I'm making a complicated app and Codex is stuck at simplifying it, so I'll give Claude a go at it.
Which models? It is not useful to compare harnesses without this information.
Which models, and at what effort level?
Summer 2026 meta (well, mine):
Sol is for routine work, Opus for frontend/design, and Fable for more complex / ambiguous / architecture work. Fable works extremely well to drive Sol as a subagent.
Fable is the only one you can actually trust to not look at the code, but Sol is somehow still more pleasant to work with, especially in fast mode. Opus is the enemy, and it will make you insane if you talk to it for too long.
When discussing Claude vs. Codex, etc. I find it necessary to make the distinction between the models and the harness.
Claude's models in my experience do a better job of inferring my intent, or to say it does a better job of giving me the result I imagined in my mind. A recent example was a UI prototype I was building for a desktop application. I had asked GPT's 5.6 Sol to update the open document in the prototype to better reflect the context of the feature I was designing, and 5.6 Sol took it very literally and had just added some text to the currently open document, not what I had in mind. I tried again with Claude Opus 5 and it added a completely new tab with a complete new document that, although imperfect, much better matched my expectations.
You could say this was a prompting skill issue, but seeing how many people are prompting their AI I believe the labs are incentivized to continue to improve their ability to infer intent.
When it comes to the desktop applications though, I find Claude Desktop's output to be incredibly verbose and full of jargon. I feel like it hits me with an entire essay and the UI doesn't have enough typographic hierarchy to make it easy to scan. ChatGPT Desktop is much better in this regard, I feel the output is concise, clear, and gives me just enough info to feel in the loop without being overwhelmed. Even though I have the setting on for technical language, it feels more understandable than Claude. I also feel that ChatGPT's desktop app has a better design and much more polish.
I do not really like how bloated both applications have become though. This weird segmentation of Chat, Work, and Code all just seems like it's pushing a technical limitation onto the user. The other day I opened a document in ChatGPT and asked it to do something, then it told me it could only do it in work "mode", so it then created an entirely new conversation with a reference to the previous conversation. It wasn't a completely new area of the UI either, it just added a "Work" badge to the new conversation in the list. Feels a bit unnecessary, like couldn't you just keep it all within the same conversation?
For me, codex $100 mo/plan and a claude teams account at work (mostly sonnet, some opus), Claude basically feels about as effective as Codex did 4-5 months ago pre-5.6. Claude still has weird patterns of being confident in one answer while another chat with the same model is confident in another answer, where one answer is clearly wrong. Missed details, over-engineering in places, while still overall helpful and effective. Codex, however, just feels freaking rock solid on Sol high. I literally have zero complaints.
guys, everything is about the harness, not the model! I even use Hy3 for free from nous portal, yes, it finishes the task in 50 minutes instead of 5, as Terra of codex would do, but I have it under the same harness (Hermes Agent) so result is more or less the same and it costed me nothing and worked at background while I was doing other things :)
I believe that “it’s about the harness” captures an important idea, but I also think it’s helpful to consider the harness and the procedure separately.
The harness provides the model with its tools, context, and environment for execution.
However, I prefer not to have the model within that harness also bear the responsibility for remembering the process steps—like plan, implement, review, fix, and verify—deciding when to move from one to the next, and keeping track of the loop's state.
For tasks that need to be repeated, I’ve been moving that part into a reliable, deterministic runtime. Inside it, Claude, Codex etc simply take on interchangeable roles.
This setup makes changing models much simpler: the overall process remains consistent, and each role can be optimized independently.
I’ve been developing this approach as ctx.traits, if you're interested, here are the docs: https://ctx.company/traits/docs/quickstart/example/
"How this article was written I wrote this article and used Grammarly to proofread and fix it."
What a brave new world we're in, where this is necessary. Regardless, it's appreciated. Although, I have the feeling that those using an LLM to do most of their writing will be less likely to include such a disclaimer.
I’ve been experimenting with this for a while, and right now I’m using Luna xhigh as my default. Previously I was using Sol medium.
Sol medium is a great balance between speed and being thorough, but it’s quite expensive. Luna xhigh seems to compensate for slightly lower intelligence by thinking and reasoning for longer, so tasks can take more time to complete. But it’s crazy cheap.
I also have some custom evals using promptfoo to make sure I’m not introducing regressions when switching models. So far, Luna xhigh has been really, really good for the price.
Don’t sleep on it. Give Luna a try.
> Changes created by Codex had fewer comments in Ruby/Ruby on Rails code. I liked that a lot, and I will soon share some experiments I ran on this.
Why is fewer comments a good thing?
Claude can add comments in code that are better just living as a comment on the PR.
You'll ask it to do something and it'll comment the code with an answer to what you asked it, rather than just explanatory comments to whoever comes after.
There's also a second issue that if the code is actually incorrect, the comment can nevertheless bolster the case for it.
Yeah, Claude insists on prompt-affirming comments, and often refers to the previous draft or implementation instead of the status quo.
> Claude can add comments in code that are better just living as a comment on the PR.
Not to Claude – its own, old comments have helped me/it solve new issues on more than one occasion.
Claude assumes humans won't look at code anymore. The comments are for the next Claude that comes in to change something.
I found that Claude would litter the codebase with what felt like notes-to-self. Sometimes it would make an abstraction that wasn't very well thought out and when I pushed back on this choice, it would usually go with something more along the lines of what I suggested in my pushback, but left a comment like "use the git CLI instead of re-implementing git" or something.
It felt like it was commenting on the diff sometimes instead of what the code was doing.
Claude comments look like this
//add returns the sum of x and y
//per section 2.1 of addition-implementation-plan.md sum is designed as the seam for user addition interfaces.
//previously sum added numbers, now it adds numbers
def add(x, y):
just reading this comments made me despise the fact that I have to read this shit again tomorrow
It's really time to move to OpenAI...
I’m trying to pitch my boss on something like bedrock/foundry/vertex where we can just swap the model and still have it serverless over an api.
Digital ocean particularly looks promising as well.
Claude is creating verbose comments with recent models and people are are tired of it
Those verbose comments provide context for why something is there way it is, so it can take those decisions in account when making changes in the future.
> why something is there way it is
It writes out stories describing what isn't there or what used to be there. It's usually not helpful, just noise. It also likes to write it in very verbose AI-styled prose.
And the way they are written is rather unhelpful: '//this not cosmetic: this css helps the user understand where he should click'. The 'this is not' pattern is useless at best.
"don't write a goddamn novel" <- I've said this to claude way too many times, nothing you can do actually seems to make it significantly less verbose
It's likely not helpful to you, but it is to Claude. Basically a memory entry at the exact point of relevance, with the same caveats. One advantage though is it survives cloning, so another using the project benefits.
It's often is not helpful to claude either because it just confuses it more most of the time.
Confusion usually surfaces immediately, so it can be steered properly. Being sure of a decision that turns out to be wrong usually means a bug that could bite when you least expect it.
It's only helpful to Claude in that it wastes more tokens and costs more.
It does visibly use a few more tokens in the short term. But it hurts Claude's ability in the long term if removed because it has to either think more every time it encounters that section in the code (more invisible thinking tokens used), or there's an increased risk of making a maladaptive change without context. Heck I've been bitten enough times already by it, though more recently it's cross project concern where Claude rediscovers some root cause that was already found and documented (I have it save a centrally-located writeup for things that took serious work to learn, when I remember, but I still need to make that knowledge pool known to other agents/sessions without having to manually prompt each time) elsewhere.
Problem is today's LLMs don't have the long term memory that humans have, and so remembering the reason behind a given change/decision has to be preserved in some way if it's non-obvious. Hence why there is {AGENTS|CLAUDE}.md, the auto-memory system, and 1001 variants of memory implementations in the wild. All are trying to ensure that LLMs can have the context they need at the location and time they need it. And you want to block Claude from using a technique that it natively finds helpful.
If they have in mind what I think of when I think of claude comments then frequently they are only relevant to the back and forth that produced the code and are exaggerated to the point of being misleading. A very common example is to land on something that isn't quite right or doesn't quite work, prompt an adjustment, and then the adjustment will have some grandiose prose about "preventing the critical bug that shipped once." People just need to clean up such litter manually but it would be nice if the models generated less of it.
The story of how the code got to a certain state belongs primarily in git commit messages, not in comments. Comments should serve to aid in understanding of the code, highlight footguns or hidden dependencies, etc.
Claude very often litters code with comments about decisions that were made within a single session/pull request, its just noise.
If one needs to dig history, then git log -p and similar are way more useful and reflect the real history rather than LLM inference about it. Comments reflecting the goal are more useful as neither history nor code itself covers it.
Looking into git history is an active decision that's usually only done after something undesirable happens, and now there's the need to try and find why. Better to have the context passively loaded as part of working with the relevant section of the code, and likely prevent that something from even happening in the first place. Think Chesterton Fence.
Those massive comments generally refer to intermediate states that never shipped, making them utterly useless, noisy and confusing as a permanent artifact. They're sometimes useful to reviewers, so a commit message or pull request comment is much more appropriate.
> useless, noisy and confusing
That's your perspective. For Claude that's an extension of its thinking, which makes it work better. Just like the person who takes notes so they have references for later. Take it away and you're negatively impacting outcomes.
No, it's the other way around. If you put a comment "don't do foo" in your code, you're biasing future readers, human or artificial, to think about foo when you explicitly want them not to.
It isn't about making readers not think about foo. It's about biasing thought about foo in a particular direction. And I've seen it work multiple times: agent mentions it encountered X comments in the code and revised its action as a result (yes, I read live agent transcripts).
Yes I hate when Claude explain in comments how it is now compared to how it was before, or how it could be if a different decision had been taken.
Dude, just talk about the current state of the code!
Claude comments often contain the whole iterative chain of decisions that led to the current state.
Useful for the LLM to know the "why", but not something a human would do, unless it's a very critical and confusing part of the code.
Fewer comments is generally a bad thing.
Fewer AI-generated comments is generally a good thing.
Good code is self-explanatory and don't need comments.
I can really recommend the book Clean Code, here is a summary: https://gist.github.com/wojteklu/73c6914cc446146b8b533c0988c...
I find claude often comments about what the code *does not* do. Relevant at the moment, but doesn't belong in the codebase.
I've been using Zed mostly with Codex session windows and its great. It is much easier to handle multiple sessions than in the Codex VSCode app and the interface is quite good.
I find Claude more often gets my intent without having to spell things out for it, while Codex gets hung up on minor details and over-engineers a solution for them.
Kinda interesting reading people comment with completely contradictory anecdotal experience with the various models.
I think a part of this is that people tend to undervalue their own skills and expertise when talking about these anecdotes.
A lot of people in the comments do have a software engineering background. People at different skill levels in different backgrounds are going to be using these tools in different ways, and that's going to heavily impact their experiences with these models.
Sure, there are differences between Fable and Sol. But I've even seen people on here saying that they're getting better mileage out of Qwen models they're self hosting.
I think the driver is just as important than the car, when it comes to this sort of stuff.
I think this definitely applies to fable/opus/sol, mixed with the undeterministic nature of the models. But those claiming Qwen are just outright coping. Thats nonsense imo
Agree with most of these.
One thing I don’t love about codex/sol is I find it tends to overengineer and be overly cautious.
I was using it to do create some scraping + data processing.
It went kind of crazy on the provenance, need at least 3 sources of consensus before promoting facts type bullshit.
defined a bunch of enums and gates.
I just wanted scrape some site data and put it into a SQLite dB. Like chill codex.
I feel like Claude is better at that.
As a follow up.
I feel like codex/sol is better at well scoped hard technical problem.
Where it can sort of run this brute force analytical loop.
Like doing performance optimization or other search type problems. I think the math proofs are good examples of this.
Agreed, only LLM I can give a folder of packet traces from a multiplayer game and wake up the next day with a server that feels identical.
I've noticed that too with Sol (xhigh) it starts with pretty good architectural goals, but really get lost in the weeds with certification and validation. Like it won't trust native tools like (in my case packer, Ansible, gcloud) to fail reliably under possible error conditions, and insists on implementing custom verification.
It also doesn't have a clear idea of what the actual threat model is, and builds all kinds of extremely defensive systems to account for imagined hostile actors. I'm like "Dude, it's only our systems that are creating these SVGs, they're never going to be user supplied, so you don't need to write an entire validation and sanitation framework here."
It also seems to treat the desired initial state of something as a permanent invariant and designs elaborate tests to ensure that it remains that way. Then when you make one little change it has to go and update a ton of tests it created.
I've had to rip out a bunch of overengineered jank from several feature implementations, and in doing so I ended up having to create retrospective documents that warn against this kind of behavior that I'll have the model review whenever a plan begins to go sideways.
“Extreme defensiveness” is a good characterization.
I wonder if it’s an artifact of OpenAI’s values or rl training approach.
Also, it prob does make it perform better just not more efficient.
Great for the OpenAI employee working on security scanning who doesn’t have to pay for their tokens.
Not so much for the dev building their web app who is trying maximize their subscription.
Like hiring an aerospace engineer to build you a shed.
What, you don't like it proposing sha256 checksums and state enums for everything?? :D It does seem to have this predilection, but I find it calms down if I tell it to simplify for "this phase". Sure, we'll pick that up down the road.. I promise. :D
I did the same this week.
The speed is the first big contrast; I have a routine multi-step skill that I run several of per week. Opus 5 was routinely taking 2 hours to do it, while older Claude models took around 20 mins; Codex restored that speed.
Second is legibility. Somebody wrote in one of the related discussions yesterday that Claude's current linguistic contortions could legitimately be considered damaging to mental health, which doesn't seem (too) hyperbolic to me. Codex (Sol) isn't perfect but it's much more direct. And so far I haven't seen it display much of an attitude, vs Opus's infuriating passive aggressive sulky know it all personality.
I slightly prefer Anthropic to OpenAI as a company, but I will vote with my wallet and discontinue my max subscription unless Anthropic does some serious damage control within the next week or two.
Claude is pretentious abrasive snob, codex is the blue collar guy that gets shit done in my experience.
I broadly agreed with the authors experience, although I wouldn't say codex does anything "wrong". I think different agents/LLM's have different personalities, and it takes learning to understand how to get them to do what you want. I had the same experience when we started using claude at work, and I was trained on codex. Claude seemed to do everything "wrong", because I was writing instructions designed for codex.
I do agree claude looks for more things to do in your repo, whereas codex is more likely to do what its old and stop. Which is better is personal preference as far as I can tell.
I honestly would've used Claude more if not for the confusing presentation/font, I'm not an expert on this subject, Codex is just a lot easier on my eyes.
Yes I have tried different settings already.
>> I think the main difference I feel between Claude and Codex is that Claude tries to go above and beyond what is asked and guess what you might want and then directly do it, while Codex is more like a companion that does what you tell it but will not overdo it. It will stop at the first sign that it might be done.
Damn, my experience is the complete opposite of this. I have posted about it a few times, e.g. https://news.ycombinator.com/item?id=49348265
tl;dr I gave GPT 5.6 a small-medium sized ticket, which should have been several hundred lines plus tests. It ended up creating a 25,000+ line diff. Another GPT 5.6 Sol with fresh context looked at the worktree and said 98% of it should be thrown away. Claude thought the same, and suggested that several dozen compactions the model went through over several hours must have caused it to go adrift. I guess that's one consequence of having a relatively small context window.
I still use Sol quite a bit. I find that it's consistently the opposite of what the author describes: it's too relentless. It doesn't know when to stop. Opus is the opposite: it'll give up a bit too easily. If everything goes well that's not an issue, but often times it'll say things like "task is done, btw I couldn't do X Y Z" and X Y Z will be some important verification step that failed because another agent was using that resource or something.
At this point I trust GPT 5.6 mostly with surgical changes, or general codebase exploration tasks. It is a faster model, so it's easier to get small things done with it. For everything else I prefer Claude, despite its annoying tendencies.
you hear things like this about every model. they're probabilistic and you can't trust them, although i doubt that kind of odd behavior would happen on low or medium effort levels. i find all of the models today pretty good at following instructions generally speaking. i'm sure you did /plan, but i rarely see it stray very far from the plan. Then again, I never walk away, since i'm always spinning up another task that can be done in parallel, and I don't let things run overnight because I'm not made of token cash (yet)
I run sol on medium about 90% of the time. Maybe that's worth a try. Especially for a small change.
I used both.
Claude Code seems more generous with its quota, which is why I use it as my main driver.
That said, Codex does seem more capable, terse, and faster. There are some tasks that Claude can't handle but Codex can. One example was a WinForms binding/project deserialization bug. Sorry, the code is a mess, so even I couldn't quite figure out which part was causing which problem.
I initially thought the bug would be difficult to reproduce in a unit-test setting. Claude could only narrow down the problem and tell me where to put a breakpoint. Codex, on the other hand, actually managed to create a reproducible unit test first, and then used that to fix the bug. That impressed me.
The only problem is the quota. Codex burns through it very, very quickly, even when I'm just using Terra 5.6 Medium. That's basically why Claude Code remains my main driver despite Codex seeming more capable.
That's surprising to me. I recently switched from Claude to Codex because I was blowing through my Claude limits. With Codex I haven't been able to use up all of my usage on the 20x plan, with Claude and fable I could do that in a day. Codex is also a lot faster than Claude. I do think Fable is slightly better and can come up with better abstractions than Codex but trying to read it's output became really frustrating to me.