Any previewers have hot takes? I've really preferred gpt-5.5 over Opus 4.8 for data analysis and scientific software work. It seems much more reliable. Fable is unusable for the type of work that I do (due to guardrails). Really looking forward to trying these new OpenAI models out.
Interesting to hear people like gpt-5.5. For me it feels smart i one shot prompts, but if you try to build up session context before doing something it feels magnitudes inferior to Claude.
I'm almost sure its because the thinking of previous turns is stripped with the responses API, so if I tell it to analyse something deeply, what remains of the understanding in future turns is only the short response text of that analysis
> thinking of previous turns is stripped with the responses API
Why do they store an encrypted reasoning payload in the session file and pass it to the API? Just a ruse? Reasoning isn’t even that many tokens, you think they’d degrade their model quality like that?
Reasoning messages would be lost immediately after a single tool call, unless you mean they sometimes go back and strip the reasoning channel retroactively, but that would increase costs via cache invalidation. I just don’t see any way it would make sense for them to do.
And wouldn’t this be noticeable by reasoning tokens not being accounted for in the context window usage?
Do you have a source for this? I'm pretty sure responses api is only there to obfuscate reasoning, but that they're still keeping reasoning traces in the backend.
> Input and output tokens from each step are carried over, while reasoning tokens are discarded.
Keeping reasoning tokens around is better for caching and for remembering past insights, so you might reasonably wonder why we designed it this way. The main benefit of dropping reasoning tokens is that you can fit a lot more work inside the model's context window before you're forced into a slow and lossy compaction step. This was a larger consideration with our earlier reasoning models that had shorter context windows (~200k), longer thinking times (up to ~100k), and poor compaction. However, now that we've shipped longer context windows, we've trained our models think much more efficiently, and we've made compaction way better than it used to be, the balance of factors is changing. Tune in Thursday!
> fit a lot more work inside the model's context window before you're forced into a slow and lossy compaction step
This is something I never understood. Why the reasoning is not included until the context is full, then the reasoning stripped optionally to allow the conversation to continue. and only then when its truly full offer a compaction. Was it to optimize caching? Well I guess it doesn't matter now that you hinted that this choice was made because of prior limitations and may change very soon
> Why the reasoning is not included until the context is full, then the reasoning stripped optionally to allow the conversation to continue. and only then when its truly full offer a compaction.
Models are typically trained (at longer conversations/more turns) either with or without the reasoning still in the conversation. If you train a model with those, then using it without them, the model will perform a lot worse, same vice-versa if you train without but then end up using the model with them.
That's why you'll see some models have it and others don't, and trying to use them another way, will make them worse, they weren't trained like that.
So why aren't the models trained with both? I'm guessing that sort of permutation in the training would lead to double the amount of training time being needed, as you know effectively will have two variants of every session you train on, with and without the reasoning.
For compiler work I found that Sol is noticably better than 5.5 (and I generally use OAI models because I like the Codex app), but Fable was still obviously better.
Better in what way? Does it follow the goals better, does the code produce have higher quality in a testable/maintainable sense or is it just closer to how you would usually program something?
I'm sorry to hear you are unable to use Fable; my partner is in the same boat and it frustrates her immensely to see what I've been able to do with it. As someone who is working with developing new linear algebra routines, Fable is so far ahead of GPT-5.5 and Opus that it's obscene. Massively better insights and far better at handling delicate corner cases without needing to mention them. I would be stunned if GPT-5.6 is at that level, but one can hope.
> It's a damn good model. Not quite as "smart" as Fable, but it is incredibly capable. Fixed all the problems I had with GPT-5.5.
> It is incredibly determined. Will run for a day without even using a /goal. It understands subagents incredibly well and is great at orchestrating. It's super pleasant in use cases like OpenClaw and Hermes Agent. It knows iOS dev incredibly well.
> It has rough edges too, but FAR fewer than 5.5 did.
> For many things, gpt-5.6-sol will become my obvious defaults.
> It is better about [following instructions] than 5.5 was. Understands intent well and hammers until it gets there. Sometimes a bit too hard.
I feel like listening to Theo about anything technical is like consulting a Labrador retriever for advice on quantum physics.
Every time I've ever seen one of his videos it's pretty clear he has very little understanding of development or engineering. I first became aware of him from his early "unit tests are a waste of time" stuff, and it seems his skillset is building a personal brand. Fair play, he's clearly talented at that, but that doesn't make his opinion on anything else worthwhile.
Sure. If his take was "100% unit test coverage is a waste of time" I think that's not unreasonable. You could make a case that the "you must write tests before you write code, every single time!" stuff is needlessly dogmatic. I also think that sometimes people focus too much on unit tests to the detriment of end to end tests that better model actual system interactions.
None of these were Theo's take. He was pushing the idea that unit tests in general were a waste of time because you could be shipping new features instead.
https://www.youtube.com/watch?v=pvBHyip4peo for an example of this. The nicest possible interpretation on this is that he's deliberately saying something he knows is wrong to attract attention.
Yep, I'm in full agreement. When extending functionality of some already existing code it also generally makes sense to write tests first.
I think the value is much lower (maybe even negative) when you're still trying to work out what shape the code will take, in an initial implementation.
Of course, as others have pointed out, nuanced opinion doesn't get clicks or YouTube views.
He doesn't believe that unit tests are complete waste of time. Just a relative waste of time. He doesn't mind AI agents writing tests. It's just mostly waste of time for humans. Because the value you get for them is not worth developer time in most cases. It's worth agent time.
When I start using a chainsaw or a car I hope it has been tested (!) Without tests before delivery the one who tests is the end user. Disaster for a unreliable chainsaw, very unpleasant for a software.
But you're right, the goal is not to write test but to ensure delivery of a reliable software. However each software is a prototype, something that has never been made before (unlike a manufactured car or chainsaw) so the customer must be ready to some unexpected behaviors when the software is released.
Since tests are often sloppy or does not cover every edge case, I see a real value for GenAI. It also forces to write good spec: very clear about inputs and the invariants for each use case. I think that AI (especially GenAI) should first be a solution to existing problem, lack of tests and good specs is often one of them.
If you write too many preconditions, postconditions, invariants etc. Then you cement your software and you will spend most of your time on the tests rather than on the actual useful software
> it's pretty clear he has very little understanding of development or engineering
I cannot prove it but I have a feeling that you may be conflating "he clearly has different opinions on things I consider non-negotiable" to "he doesn't know what he's talking about".
I also watched a lot of his videos. I wildly disagree with him a lot of times, but he has his reasoning, and I can see (and verify!) that those ideas are coming from an engineering perspective.
He's clearly very knowledgeable about some things, but I think he has harmed his credibility be becoming a 'tuber who prioritizes thumbnails and hot takes over engineering.
Not really. If you're a YouTuber it's necessary to follow the algorithm which includes making such a YouTuber face, clickbait actually works and has a direct financial correlation as Linus Tech Tips has shown.
If you are a competent engineer already why do you need to create self demeaning clickbait content on YouTube? Narcissism? I aggressively block any such content, because if you click on any of it, you easily get sent down the YouTube spiral of crap.
One can be multiple things. That you find it demeaning is your personal opinion and frankly is more a reflection on yourself. There are many like Casey Muratori that are competent engineers as well as YouTubers.
Casey (on his channels) does not do the clickbait thing, which makes all the difference. You can either maintain integrity and credibility or you can make brainrot shockface “it is over” content. There is no middle ground.
Maybe you are desensitized to it, but I have a carefully curated YouTube and I know that it can be a completely different platform experience if you reject with prejudice any such content.
Use DeArrow if you're so sensitive to it. And like I said, it is necessary if YouTube is also a money making endeavor. Casey likely makes little from his videos so he doesn't need to have clickbait thumbnails.
DeArrow gives a YouTube page with rageslop that has been marked as rageslop. I do not need an addon to determine that, I need a feed that does not serve it to me in the first place, and that means curating my feed by rejecting certain videos and channels.
And this goes back to my point, if you are a competent engineer, why are you spending time producing rageslop for money, rather than, you know, doing the engineering.
There's also money in being an engineer. If he makes so much more money as YouTube rage farmer, then I'd be surprised, but it still wouldn't change my perception of it being degrading work, harmful to his credibility and probably harmful to his mental health too.
I sort of disagree, the issue is that he like so many professionals (prime agent, being the other) becoming youtubers uses their experiences to make their opinion the only opinion when said opinion is nuanced or plain wrong objectively.
No he usually acknowledges other opinions (including the ones that I share) and tells from his perspective why they are wrong. It feels condescending when you see a face on screen, roasting what you think is right, but I personally could get over it and learn to take the bits that challenge my ideas.
Of course, youtube isn't interactive and when you see something that you think is objectively wrong, your options are writing a comment nobody will read or ignoring it, which is frustrating, but that, in my opinion, doesn't discredit the content producer itself.
You can make a counter argument video, which I've seen done, and then sometimes the original YouTuber will reply or collab. This happened when Theo had some truly awful takes on Flutter while knowing nothing about it, a prominent Flutter dev replied rebutting each point made, then they both collaborated to make a video where they switch technologies for a few hours to build something, and it was enlightening for both parties.
The problem is that he's a tech influencer first, a tech expert second.
That's his motivation, influencing. Not teaching.
I'm not so against him as the previous commentor but I feel basically all YouTubers who have succeeded in building a brand have the same problem. They have to present their opinions as unassailable truth, they can't allow nuance. Within reason of course, they also have to play the game of appearing considerate and understanding of other people but it will always boil down to proving they are the real experts, their ultimate goal is always to get you watching more of their content.
If they don't do this, they appear less trustworthy and their brand wouldn't have grown as much as it did. They might genuinely have some expertise to share, and even contrary or downright wrong takes could teach us something if they were only presented in a way that encourages critical thinking. But as the person above said, their real deep expertise is in brand building.
It's curious how so many people get triggered by a smart person saying what he believes to be true. Yes, he is pretty amazing. Yes, he is rarely wrong. No, it doesn't affect you or me in any way because he is not in competition with you. Go do something else if you don't enjoy his takes.
I don't get many programmer influencers in my feed that deal with newsworthy relevant stuff. Theo is the least wrong and most humble one in my perception.
I disagree with this assessment. Theo is nerdy and, yes, he has a healthy ego. But, he provides insightful commentary on his channel and he works very hard to present what he believes is the truth. Compared to many YouTubers, whose content is vacuous, Theo is mostly the real deal.
He has big 'theatre kid' energy (at least certainly had, watched him years ago) - he desperately wants to make clear that there's a group of cool kids and he's in it.
His youtube channel used to be about talking about the new FOTM Javascript framework/technology - not presented as 'here's a cool thing, let's check it out' but 'everyone worth a damn already uses this, get with the times grandpa'
I wouldn’t call it a recipe for disaster, but oh boy if you leave an agent that “hammers until it gets there” on its own with an underlying bug in a dependency…
For some tasks, there is no amount of "steering" that will produce sensible code. The model needs to be sufficiently capable as a baseline; this is the "intent" that people are referring to with Fable.
That doesn’t sound like the “it hammers until it’s done”-type of intent.
Just last night Fable decided to get into a rabbit hole of debugging a database driver issue by packet sniffing the network traffic instead of just adding debug statements to the code. Definitely needed steering, and I don’t know many people whose first intuition would be to use pcap when they have a segfault.
Overengineering is the name of the game with Fable. Sometimes you don't want that, sometimes you really do, especially as a researcher. It's a very nice tool to have around for those special tasks.
> Not quite as "smart" as Fable, but it is incredibly capable.
THIS IS BECAUSE GPT-5.6 SOL IS... just a more posttrained version of GPT-5.5, not a brand new bigger model than GPT-5.5. It's not like how Mythos is bigger than Opus.
OpenAI switching to Sol/Terra/Luna renaming is just a way to rip off people and charge more money for the same sized model.
GPT-5.6 --------> GPT-5.6 Sol
GPT-5.6-mini ---> GPT-5.6 Terra
GPT-5.6-nano ---> GPT-5.6 Luna
Except OpenAI is about to charge way more money for GPT-5.6 Sol and GPT-5.6 Terra, than if they named it GPT-5.6 and GPT-5.6-mini.
My feeling is that GPT-5.5 doesn't lack the raw intelligence so much as it lacks "methodology". I don't know how exactly to put it... how to approach a problem, how to take care of the details and side effects, how to handle unexpected difficulties and bugs, how to not spin out of control, how to write solid code, how to clean up afterwards, how to document, how to give useful feedback... the things that you learn on the job.
So, if they improved a lot in those areas, then GPT-5.6 could become a lot more useful compared to GPT-5.5 even though it might score lower in many benchmarks. It's possible but unlikely since their approach was mostly brute force in the past.
Is Fable really that much different? I almost instinctively create elaborate processes, workflows, set up a bunch of linters and dump research docs any time I bootstrap a new project regardless of what model I'm using. They all spiral out of control if they're not following a predefined process.
Very. Fable 5 is incredibly efficient token wise, second only to GPT-5.5 and is far more affordable run-to-run than the pure input/ouput costs would suggest. Task adherence, task inference, tool calling and task assessment are all significantly ahead of GPT-5.5, especially as the later strongly degrades the second compaction comes into the mix, I suspect because of OpenAIs obsessive optimisation of reasoning tokens into a hard to read (and thus also hard to compact) mess.
Fable 5 meanwhile has a reliable 1m context window and compaction that the few times I did eval it does also do well. Not quite as easy to trust as GPT-5.4, but that's mainly because with thats 272k context window I simply got more familiar with GPT-5.4s incredibly dependable compaction.
Purely concerning encoded information wise, Fable 5 is near or on the same level as Gemini in my limited test set focused on those tasks, which in very niche cases can make a difference even with coding, but the truest advantage for coding assistance (besides frontend/UX) is that the code Anthropic models provide is more parsable. Hard to explain, but I can read, follow and mentally map Fable 5 (and even Opus 4.5-4.8) output far more than GPT-5.4 or GPT-5.5 code.
Task orchestration and (more importantly) knowing when to recommend against using such vs Opus 4.8 is another strength of Fable 5 I've use liberally, computer use is also solid, albeit not as token efficient as GPT-5.5 for my limited use cases.
That being said, we are still far from a stage where I'd not want to review the output, but yes, I do rate Fable 5 very highly. GPT-5.5 can have a similar ceiling but long horizon has become less usable over GPT-5.4 and in either case, parsing their output is (far more) of a chore. Maybe post training can address some of this, hopeful on the compaction front myself. Also interested in what happened to OpenAI models on AWS Trainium, I was expecting that to be a major boon for their commercial adoption, but haven't heard anything since then...
Yes it is. With Fable you don't need to create any sort of elaborate process, it seems to understand the user's intent much better than Opus class models.
I use Open AI and Claude a lot right like a lot everyday for hours multiple hours. Open AI gives much more value for money than Claude much more I'd say x 10. Mainly I use it for writing fiction books and literally Claude is locked 90% of everyday trying to jip me for tokens. It's not as good at coding for what I do which is a very complicated application. However it is very good at writing it's really good which is why I keep it right but over 90% maybe actually all of my work except the initial draft of a chapter is done by open AI.
No, GPTCyber is specifically trained for cybersecurity, and GPT-5.5-pro is just an ensemble of many subagents, not an actual model.
Mythos is simply a much bigger model in terms of parameters and I don't think OpenAI will have anything of its size anytime soon (My theory is that OpenAI had given up on scaling up parameters after GPT4.5 flopped).
Post-training can have big gains no? I don't think the current sizes at ~1T are saturated in intelligence (it's like saying AlphaGo Master is just a post-trained version of AlphaGo Lee)
I looked at his YouTube, and found a stream of industry gossip and beginner content like "web dev tutorials". I have nothing against such content and it may be useful and good fun to watch.
But does that say anything about this particular model? People have been using models effectively for web code since Gpt 3.x.
I will stop here, sorry but I think we have limited time to listen to opinions and nowadays since they are abundant on social media we should give preference to the substantiated ones.
I’m bouncing back between Codex and Claude like a ping-pong ball. I much prefer the experience using Codex, less verbose and to-the-point I’ve found. But Fable, being as strong as it is, is a big draw for Claude right now. I’ll likely switch back to Codex if 5.6 Sol is comparable.
Same. For some reason late opus model are very superficial doing ux work and so am using gpt for that, but backend is much better engineered by claude, gpt prefer to duplicate everything it needs on the spot causing class sprawl
How are y'all carrying context history from one agent to the other?
I also flip between the models due to quota, TUI enhancements, model updates and service availability.
To handle this, I built a thing that normalizes your transcripts between Claude Code and Codex into a shared DB, then a CLI and skill.
It has made it so it doesn't matter what I built where (or when) I just refer to the work and drop in a /total-recall (or $total-recall on codex) and the agent brings it into the current convo.
I realize there are a lot of ~memory tools out there, but I think particular my approach and product behavior is unique.
Personally I just , in the orchestration loop, have all decisions be constantly reviewed and deliberated on and the decisions logged in a permanent way, that way everything is auditable, the model if needed can go back and look at why x or why decision was made or x or y tool used, and they're all labeled as D-1234 or whatever.
Plus I have it log the council discussions and always include provenance or the opinions so fable can go back after every major implementation and review how the orchestration loop could be improved. Basically have it log as much thinking in an organized compartmentalized way is better than any memory feature I've found though I haven't tried many. Auditable logs for every major decision, use 5.5 with reckless abandon (still have 3 resets myself).
Not claiming this is perfect but it has led to a very easy time of any fresh agent picking up the project. I also keep a task queue and project status and agent playbook that also get refined based off the logs of how a run went
How can be locked? If you have a proper agents in your project it will work out of the box with any model. I use codex and Hermes on same project with 0 issues. Skills, MCP and other features are useless imo.
My agent has access to glab with a user and can do whatever within permissions. No need a MCP. MCP maybe just for browser control.
5.6 Sol is extremely good, definitely Fable level from my experience. With 5.6 Sol being half the price and noticeably faster I think Anthropic will find the coming months unpleasant.
I would be absolutely stunned if this were really the case in general given how irresponsibly large Fable is, and 5.6 Sol most definitely is not. It depends on what your problems are though, I suppose, since there are those that swear Fable is at best a minor upgrade over Opus, which has not been my experience.
I personally like Codex much more as vscode integration. It's actually nice to use, looks great, handles images much, much better, can even send me images in the chat to see what it's doing(my side project is based heavily on image processing), and what's most important to me is that /steer actually works. When I type something to Claude mid-task, it'll maybe acknowledge that within next few minutes, although it seems to be actually quicker last few days but still takes a minute or two, whereas Codex will almost instantly read it, switch what it's doing or answer me. It feels much more polished in some ways(though usage page in settings almost never loads for me).
Damn this is exciting. I love that gpt models are much faster, efficient and cheaper than Claude models. They are so fast even on high/xhigh that I don’t find myself using the parallel agent setup anymore much since its cognitively less demanding to just follow along what the model is doing and most tasks it will complete in <5-<10mins anyway.
THIS IS BECAUSE GPT-5.6 SOL IS... just a more posttrained version of GPT-5.5, not a bigger model than GPT-5.5. It's not like how Mythos is bigger than Opus.
OpenAI switching to the Sol/Terra/Luna naming is just a way to rip off people and charge more money for the same sized model.
GPT-5.6 --------> GPT-5.6 Sol
GPT-5.6-mini ---> GPT-5.6 Terra
GPT-5.6-nano ---> GPT-5.6 Luna
Except OpenAI is about to charge way more money for GPT-5.6 Sol and GPT-5.6 Terra, than if they named it GPT-5.6 and GPT-5.6-mini.
Two important things to note, if you want to verify what I say/correct me:
GPT-5.6 Terra actually scores worse than GPT-5.5 on many benchmarks. It's not GPT-5.5 trained with more compute; it's basically GPT-5.6-mini that's been distilled from GPT-5.6 full size. Remember, GPT-5.4-mini had almost the same benchmarks as GPT-5.2 after all.
Opus 4.8 runs at ~90 tokens per second. Fable 5 runs at ~40 tokens per second on from Anthropic, because it's a bigger/slower model.
A few days after the release, when the dust dies down, look at how many tokens/second GPT-5.6 Sol is running at. I will bet it's the about same as GPT-5.5, and not half the speed. (OpenAI is not incentivized to slow down the model for paying customers). But the model tokens/sec will be a big clue- if OpenAI is charging more money for the same sized model or not.
not just that, but the entire industry spend several years seeking investment on the "pure" idea that they just need more compute and more parameters to reach AGI.
And the "business" obvious is still doing that but the science and implementation has be realizing that this just isn't true. They're not getting AGI out of a single LLM by itself.
right, but the business arm will always be dominant. I see what Chinese models are doing as the same as japanese car models in the 80s: producing smaller, more efficient products that address the realities of the "total addressable market" that no business model would support. They're, unfortunately, providing public value where the US and Europe used to tread.
There's a lot to make efficient, but it should be clear to everyone that just throwing compute at larger models isn't going to magically make it rain.
I'm most curious about whether OpenAI finally taught its models how to design interfaces. They have been behind the other labs in this area for what feels like ages.
Probably, but I think it's too little too late. Not much point to it if it's not permanent. The "get the most out of Fable until it goes away" frenzy is getting old fast. The cybersecurity blocks are very obnoxious too.
If OpenAI can launch a Fable tier model that's actually usable on a subscription, then Anthropic is just going to lose, and badly.
Agreed, this is one of the things I'm very surprised - one would think that a product like this is managed more consistently, but every few days there is another announcement or change in what the subscription can and can't do and to what extent.
Same also for the announced changes around `claude -p` and Agent SDK use that were backtracked
It's because Anthropic doesn't have capacity while OpenAI does. People clowned Altman a couple years ago because of the massive data center build out commitments but that has proved to be quite prescient. It is why Codex has much higher, almost unlimited limits, while Claude Code rate limits hourly and weekly much more.
I find codex way more usable. It’s not pretentiously verbose like Claude. It’s also responsive - I can see the progress easily and steer the conversation. With Claude, it might take 15 minutes and I would lose patience.
Both are verbose in their own way, and both - terrible. Claude models love to throw huge blobs of text in architecture planning / interview conversations, but in not a mentally draining language. OpenAI models are more compact, but very dense & formal - they will speak in RFC language for a button that clicks and submits a form.
I've seen this with GPT, and I usually ask it to put together a more easy to understand document for a specific target audience or reading level and it seems to do okay.
Coding with AI it feels like if you're not using the best model then you're possibly missing out - creating less capable, maintainable, just plain 'good' code. Why waste time using anything less than the best and cleaning up the mess later on. This is why I feel like local models and Chinese models aren't taking off (and Gemini/Grok) - they work, but they're plain just not as good as OpenAI/Anthropic. If you have the money then it doesn't make sense to code with anything else.
There are diminishing returns, especially for more mundane tasks. Fable is nice, and I bet Sol is also nice. But there really isn't much of a difference right now when using something beyond Opus or presumably Terra for most things. They're most useful when doing greenfield, highly complex/novel tasks. When Open Source catches up, it will be more widely adopted.
Yea, but by the time open source catches up, the frontier will be that much more capable and you won't want to waste time babysitting less capable models.
Another dimension for the fronteir to move in is speed. Codex has /fast which is great, but yea the bottleneck right now in many cases is just the time it takes these tools to complete tasks. I'm running many sessions in parallel just because I'm waiting for tasks to finish. I'm constantly round robin'ing them, and kicking them off on the next 20 minute task. If these models were faster I wouldn't need to context switch as much.
That depends entirely on how you're using AI. If you're getting it to do all the hard thinking, then sure using the best model is probably always going to be better. But it's also going to be expensive.
Using cheaper models and using your skills and expertise from the pre-AI era can get you working just as fast. You've gotta be more specific about the work you need doing. It's less "vibes" based, but they're still effective.
Also, Chinese models absolutely are taking off. I used Claude and GPT at work, and then I tried using some Chinese models for personal projects. I am 100% convinced they're like 90% as good for 10% of the cost. But you've basically gotta be a good developer first and know what you want and know when it's giving you shit.
Honestly I'm on $200 a month for Claude Max and $100 a month for Codex, and it's nothing compared to the productivity gains if you're programming professionally. 10 bucks a day, I spend more for lunch. Time is money and I'm not going to waste time with a lesser model if I don't have to.
Yeah if you're a professional engineer it's a no brainer to buy these subs, even multiple subs, and you could replace another employee's salary especially if you're a solo founder working on your own product.
Is it cheaper than Codex for example? The problem with paying per token via API is it's not subsidized like subscriptions are, maybe Z AI one is though. But GLM doesn't have vision which is a deal breaker for many frontend or full stack tasks.
Or maybe you are still iterating on a plan or spec file with Qwen 3.6 27B, while I implemented three features with Fable and QA tested them in the testing environment.
Of course, if you think that this approach is as fast and effective as "vibe coding" as in outsourcing more thinking to the AI, it is not surprising you would conclude the cheaper models were nearly as useful.
I don't know if you are right or not, a lot depends on the constraints of the project and team.
Some of the newer arguably now viable use cases, such as porting a large codebase to Rust, are certainly not going to be as fast with a more manual approach.
70s thru 90s computing and even into the early 2000s every new bit of computer meant new capabilities.
Eventually it plateaued and now you can do a decent chunk of your computing on something from 2012.
People keep saying scaling will top out, for example. But scaling keeps stubbornly refusing. New techniques keep coming along too. It's really still exploding into existence and every new generation brings new capability. Eventually it'll clear a ceiling for your key use cases and you'll stop worrying about new models.
It always pays to look back at history and see if you can pattern match.
I’ve been using mostly deepseek v4, kimi k2.6, and gpt 5.3-codex
I sometimes chuck a few tokens to gpt 5.5 and opus 4.8 and they can sometimes solve a problem one of the other models couldn’t, but they’re not like 10x better or anything in my experience. More like 1.2x better
Earlier I predicted that Fable and Sol would be of similar capability, I think I will be wrong. Here is why: there is no indication that there are any classifiers like in Fable. I think OpenAI found out how to lobotomise the model without classifiers but the tradeoff is that it is a weaker model. I wonder how people feel about that. Would you like a highly intelligent jagged model with classifiers or slightly less intelligent smooth model without classifiers?
Based on the pricing I guess GPT 5.6 is the same size as GPT 5.5.
I would not be surprised if it is not as intelligent as the Mythos class models.
I have seen rumors that GPT 6 may release before September. The same person also claimed that a Fable 5.1 checkpoint has been completed a few weeks ago.
I know a few of my comments are related to this, but these new names are horrible. Why introduce ANOTHER layer of confusion and drop the mini, nano suffixes that people got used to?
How does this go through so many layers of management at a trillion dollar company without who has a say raising this? I simply can't believe how stupid the naming scheme from OpenAI was and continues to be even after they acknowledged it earlier.
Honestly, "mini" and "nano" to me just seem like really awful names from a marketing perspective - they might as well call it "lobotomized crap version of GPT" and "even more lobotomized crap version of GPT".
Whereas Sol/Luna/Terra reads more like "GPT for hard/medium/basic problems".
I disagree. I needed a small text to json model that would parse basic info into a json, nothing else. I instantly knew to start with nano and if not good enough, use mini. This was obvious to me, just looking at the name. Now you have to actually know what those names mean. And I can guarantee you they'll add 5 more to create more hype, or make new names in whatever the next-gen-world-breaking-dangerous-model will be.
My theory is that they don't have Fable-class intelligence so they needed different hype vehicle :) This rename helps build excitement a bit more than just releasing ordinary GPT-5.6 increment.
Yet, in a month we'll be fine. We were fine with Anthropic naming models by music. I'm sure celestial bodies will be OK too. Larger = better. It's simple. As for the why? Marketing, making products feel "fresh", exciting, new, something alluring that we didn't have before. So, much like since industrialization.
What surprises me is not this, but that OpenAI changed things up without syncing with a GPT 6.
I mostly used GPT-Venti for the complex part, but the documentation was done by either GPT-Grande or GPT-Tall.
On a more serious note, I can vividly imagine how difficult it would be to find a set of words that could plausibly suggest a relational meaning while remaining non-diminutive in every single model name. Adding to the complexity, it is going to be used globally, and the main competitor already has a successful and fabulous naming scheme.
Nano and mini, which is smaller? This is a bit more clear imo. It also helps make reduce the expectation that bigger isn't necessarily better for all use cases.
(Well given the limited amount of things we can deduce from a name)
I've been running a custom enterprise agent on 5.4 and it's been very good so far. I am looking forward to trying it with the monster model to see if we can approach some additional business cases.
I think if you are not seeing reasonable performance in your agent loops as of 5.5, it's likely there is a deficit with how the loop, prompt or tools interact with the environment.
Fable 5 for the planning, thinking, reasoning part, then GPT 5.5 to implement is an almost perfect combo, with Fable then reviewing GPT's code.
Codex CLI just seems faster at coding than Claude Code but Fable is just a level above intelligence wise, it's truly like taking to very very very smart human.
With GPT 5.6 though will be interesting to see if things flip, to have Codex speed (or faster) with Fable level intelligence is a game changer.
I would be really interested in real life throughput. For an agentic chat situation, we are still on 5.4 - not because of the cost, but it's simply much faster than 5.5 with comparable results. Also we are using gpt-5.4-mini a lot for quick summaries, tldrs etc.
In an ideal world we would upgrade 5.4 to 5.6 terra and 5.4 mini to 5.4 luna. But does somebody already have some measurements at least in terms of speed?
174 comments
[ 0.19 ms ] story [ 41.4 ms ] threadWhy do they store an encrypted reasoning payload in the session file and pass it to the API? Just a ruse? Reasoning isn’t even that many tokens, you think they’d degrade their model quality like that?
Reasoning messages would be lost immediately after a single tool call, unless you mean they sometimes go back and strip the reasoning channel retroactively, but that would increase costs via cache invalidation. I just don’t see any way it would make sense for them to do.
And wouldn’t this be noticeable by reasoning tokens not being accounted for in the context window usage?
Our docs show a diagram here:
https://developers.openai.com/api/docs/guides/reasoning
> Input and output tokens from each step are carried over, while reasoning tokens are discarded.
Keeping reasoning tokens around is better for caching and for remembering past insights, so you might reasonably wonder why we designed it this way. The main benefit of dropping reasoning tokens is that you can fit a lot more work inside the model's context window before you're forced into a slow and lossy compaction step. This was a larger consideration with our earlier reasoning models that had shorter context windows (~200k), longer thinking times (up to ~100k), and poor compaction. However, now that we've shipped longer context windows, we've trained our models think much more efficiently, and we've made compaction way better than it used to be, the balance of factors is changing. Tune in Thursday!
This is something I never understood. Why the reasoning is not included until the context is full, then the reasoning stripped optionally to allow the conversation to continue. and only then when its truly full offer a compaction. Was it to optimize caching? Well I guess it doesn't matter now that you hinted that this choice was made because of prior limitations and may change very soon
Models are typically trained (at longer conversations/more turns) either with or without the reasoning still in the conversation. If you train a model with those, then using it without them, the model will perform a lot worse, same vice-versa if you train without but then end up using the model with them.
That's why you'll see some models have it and others don't, and trying to use them another way, will make them worse, they weren't trained like that.
So why aren't the models trained with both? I'm guessing that sort of permutation in the training would lead to double the amount of training time being needed, as you know effectively will have two variants of every session you train on, with and without the reasoning.
> It's a damn good model. Not quite as "smart" as Fable, but it is incredibly capable. Fixed all the problems I had with GPT-5.5.
> It is incredibly determined. Will run for a day without even using a /goal. It understands subagents incredibly well and is great at orchestrating. It's super pleasant in use cases like OpenClaw and Hermes Agent. It knows iOS dev incredibly well.
> It has rough edges too, but FAR fewer than 5.5 did.
> For many things, gpt-5.6-sol will become my obvious defaults.
> It is better about [following instructions] than 5.5 was. Understands intent well and hammers until it gets there. Sometimes a bit too hard.
[^0]: https://nitter.net/theo/status/2074708892341481755
(A little toning down of the goblin fetish would be nice too, haha.)
Every time I've ever seen one of his videos it's pretty clear he has very little understanding of development or engineering. I first became aware of him from his early "unit tests are a waste of time" stuff, and it seems his skillset is building a personal brand. Fair play, he's clearly talented at that, but that doesn't make his opinion on anything else worthwhile.
None of these were Theo's take. He was pushing the idea that unit tests in general were a waste of time because you could be shipping new features instead.
https://www.youtube.com/watch?v=pvBHyip4peo for an example of this. The nicest possible interpretation on this is that he's deliberately saying something he knows is wrong to attract attention.
I think the value is much lower (maybe even negative) when you're still trying to work out what shape the code will take, in an initial implementation.
Of course, as others have pointed out, nuanced opinion doesn't get clicks or YouTube views.
But you're right, the goal is not to write test but to ensure delivery of a reliable software. However each software is a prototype, something that has never been made before (unlike a manufactured car or chainsaw) so the customer must be ready to some unexpected behaviors when the software is released.
Since tests are often sloppy or does not cover every edge case, I see a real value for GenAI. It also forces to write good spec: very clear about inputs and the invariants for each use case. I think that AI (especially GenAI) should first be a solution to existing problem, lack of tests and good specs is often one of them.
I cannot prove it but I have a feeling that you may be conflating "he clearly has different opinions on things I consider non-negotiable" to "he doesn't know what he's talking about".
I also watched a lot of his videos. I wildly disagree with him a lot of times, but he has his reasoning, and I can see (and verify!) that those ideas are coming from an engineering perspective.
Maybe you are desensitized to it, but I have a carefully curated YouTube and I know that it can be a completely different platform experience if you reject with prejudice any such content.
And this goes back to my point, if you are a competent engineer, why are you spending time producing rageslop for money, rather than, you know, doing the engineering.
Of course, youtube isn't interactive and when you see something that you think is objectively wrong, your options are writing a comment nobody will read or ignoring it, which is frustrating, but that, in my opinion, doesn't discredit the content producer itself.
That's his motivation, influencing. Not teaching.
I'm not so against him as the previous commentor but I feel basically all YouTubers who have succeeded in building a brand have the same problem. They have to present their opinions as unassailable truth, they can't allow nuance. Within reason of course, they also have to play the game of appearing considerate and understanding of other people but it will always boil down to proving they are the real experts, their ultimate goal is always to get you watching more of their content.
If they don't do this, they appear less trustworthy and their brand wouldn't have grown as much as it did. They might genuinely have some expertise to share, and even contrary or downright wrong takes could teach us something if they were only presented in a way that encourages critical thinking. But as the person above said, their real deep expertise is in brand building.
I don't get many programmer influencers in my feed that deal with newsworthy relevant stuff. Theo is the least wrong and most humble one in my perception.
His youtube channel used to be about talking about the new FOTM Javascript framework/technology - not presented as 'here's a cool thing, let's check it out' but 'everyone worth a damn already uses this, get with the times grandpa'
It's shocking how many accurate tropes this hits.
If there’s anything I learned over the past 12-18 months is that this is a recipe for disaster, except for throwaway stuff.
I thought most senior engineers settled on the fact that steering a model yields much better results?
In my experience even Fable still requires guidance (although the options it provides are generally better).
Just last night Fable decided to get into a rabbit hole of debugging a database driver issue by packet sniffing the network traffic instead of just adding debug statements to the code. Definitely needed steering, and I don’t know many people whose first intuition would be to use pcap when they have a segfault.
THIS IS BECAUSE GPT-5.6 SOL IS... just a more posttrained version of GPT-5.5, not a brand new bigger model than GPT-5.5. It's not like how Mythos is bigger than Opus.
OpenAI switching to Sol/Terra/Luna renaming is just a way to rip off people and charge more money for the same sized model.
GPT-5.6 --------> GPT-5.6 Sol
GPT-5.6-mini ---> GPT-5.6 Terra
GPT-5.6-nano ---> GPT-5.6 Luna
Except OpenAI is about to charge way more money for GPT-5.6 Sol and GPT-5.6 Terra, than if they named it GPT-5.6 and GPT-5.6-mini.
Excuse me, but what are you on about?
Unless I'm mistaken, they have literally(1) stated that it will cost $5 per 1M tokens in, and $30 for 1M output tokens. The same as GPT-5.5.
[1] https://openai.com/index/previewing-gpt-5-6-sol/
So, if they improved a lot in those areas, then GPT-5.6 could become a lot more useful compared to GPT-5.5 even though it might score lower in many benchmarks. It's possible but unlikely since their approach was mostly brute force in the past.
Fable 5 meanwhile has a reliable 1m context window and compaction that the few times I did eval it does also do well. Not quite as easy to trust as GPT-5.4, but that's mainly because with thats 272k context window I simply got more familiar with GPT-5.4s incredibly dependable compaction.
Purely concerning encoded information wise, Fable 5 is near or on the same level as Gemini in my limited test set focused on those tasks, which in very niche cases can make a difference even with coding, but the truest advantage for coding assistance (besides frontend/UX) is that the code Anthropic models provide is more parsable. Hard to explain, but I can read, follow and mentally map Fable 5 (and even Opus 4.5-4.8) output far more than GPT-5.4 or GPT-5.5 code.
Task orchestration and (more importantly) knowing when to recommend against using such vs Opus 4.8 is another strength of Fable 5 I've use liberally, computer use is also solid, albeit not as token efficient as GPT-5.5 for my limited use cases.
That being said, we are still far from a stage where I'd not want to review the output, but yes, I do rate Fable 5 very highly. GPT-5.5 can have a similar ceiling but long horizon has become less usable over GPT-5.4 and in either case, parsing their output is (far more) of a chore. Maybe post training can address some of this, hopeful on the compaction front myself. Also interested in what happened to OpenAI models on AWS Trainium, I was expecting that to be a major boon for their commercial adoption, but haven't heard anything since then...
Mythos is simply a much bigger model in terms of parameters and I don't think OpenAI will have anything of its size anytime soon (My theory is that OpenAI had given up on scaling up parameters after GPT4.5 flopped).
Spanish: Sol, tierra, luna
Italian: Sole, terra, luna
Catalan: Sol, terra, lluna
Portuguese: Sol, terra, lua
Might as well call it gelatto, siesta, fiesta if they think it sounds cool.
https://x.com/theo/status/2074708892341481755
5.6 sol seems to hit a lot of the gaps with 5.5
sucks its not "mythos" but i will take it
I looked at his YouTube, and found a stream of industry gossip and beginner content like "web dev tutorials". I have nothing against such content and it may be useful and good fun to watch.
But does that say anything about this particular model? People have been using models effectively for web code since Gpt 3.x.
I will stop here, sorry but I think we have limited time to listen to opinions and nowadays since they are abundant on social media we should give preference to the substantiated ones.
I also flip between the models due to quota, TUI enhancements, model updates and service availability.
To handle this, I built a thing that normalizes your transcripts between Claude Code and Codex into a shared DB, then a CLI and skill.
It has made it so it doesn't matter what I built where (or when) I just refer to the work and drop in a /total-recall (or $total-recall on codex) and the agent brings it into the current convo.
I realize there are a lot of ~memory tools out there, but I think particular my approach and product behavior is unique.
If you're open to giving it a try, I'd appreciate any feedback: https://contextify.sh recent show hn: https://news.ycombinator.com/item?id=48777790
Personally I just , in the orchestration loop, have all decisions be constantly reviewed and deliberated on and the decisions logged in a permanent way, that way everything is auditable, the model if needed can go back and look at why x or why decision was made or x or y tool used, and they're all labeled as D-1234 or whatever.
Plus I have it log the council discussions and always include provenance or the opinions so fable can go back after every major implementation and review how the orchestration loop could be improved. Basically have it log as much thinking in an organized compartmentalized way is better than any memory feature I've found though I haven't tried many. Auditable logs for every major decision, use 5.5 with reckless abandon (still have 3 resets myself).
Not claiming this is perfect but it has led to a very easy time of any fresh agent picking up the project. I also keep a task queue and project status and agent playbook that also get refined based off the logs of how a run went
You are the sole owner of the project implementation.
User maintained documentation:
- goals.md for the project overall goals
- tech.md for guidance on how to build the project
Agent maintained documentation, current state living specs, these are not logs:
- project.md is a map of the code, components and features.
- choices.md write here all decision taken by the user.
Do not duplicate information between these document.
e.g., it still doesn't have /revise or /undo!
My agent has access to glab with a user and can do whatever within permissions. No need a MCP. MCP maybe just for browser control.
I canceled Claude plan a few months back and have been using this. OpenAI plans are much more generous.
What type of work is this for in your experience?
OpenAI switching to the Sol/Terra/Luna naming is just a way to rip off people and charge more money for the same sized model.
GPT-5.6 --------> GPT-5.6 Sol
GPT-5.6-mini ---> GPT-5.6 Terra
GPT-5.6-nano ---> GPT-5.6 Luna
Except OpenAI is about to charge way more money for GPT-5.6 Sol and GPT-5.6 Terra, than if they named it GPT-5.6 and GPT-5.6-mini.
Two important things to note, if you want to verify what I say/correct me:
GPT-5.6 Terra actually scores worse than GPT-5.5 on many benchmarks. It's not GPT-5.5 trained with more compute; it's basically GPT-5.6-mini that's been distilled from GPT-5.6 full size. Remember, GPT-5.4-mini had almost the same benchmarks as GPT-5.2 after all.
Opus 4.8 runs at ~90 tokens per second. Fable 5 runs at ~40 tokens per second on from Anthropic, because it's a bigger/slower model. A few days after the release, when the dust dies down, look at how many tokens/second GPT-5.6 Sol is running at. I will bet it's the about same as GPT-5.5, and not half the speed. (OpenAI is not incentivized to slow down the model for paying customers). But the model tokens/sec will be a big clue- if OpenAI is charging more money for the same sized model or not.
And the "business" obvious is still doing that but the science and implementation has be realizing that this just isn't true. They're not getting AGI out of a single LLM by itself.
There's a lot to make efficient, but it should be clear to everyone that just throwing compute at larger models isn't going to magically make it rain.
What is this very confident assumption based on?
If OpenAI can launch a Fable tier model that's actually usable on a subscription, then Anthropic is just going to lose, and badly.
Same also for the announced changes around `claude -p` and Agent SDK use that were backtracked
So claude: 10 paragraphs of prose
codex: 1 paragraph of jargon over jargon.
Another dimension for the fronteir to move in is speed. Codex has /fast which is great, but yea the bottleneck right now in many cases is just the time it takes these tools to complete tasks. I'm running many sessions in parallel just because I'm waiting for tasks to finish. I'm constantly round robin'ing them, and kicking them off on the next 20 minute task. If these models were faster I wouldn't need to context switch as much.
Using cheaper models and using your skills and expertise from the pre-AI era can get you working just as fast. You've gotta be more specific about the work you need doing. It's less "vibes" based, but they're still effective.
Also, Chinese models absolutely are taking off. I used Claude and GPT at work, and then I tried using some Chinese models for personal projects. I am 100% convinced they're like 90% as good for 10% of the cost. But you've basically gotta be a good developer first and know what you want and know when it's giving you shit.
I’d describe it as something between Sonnet and Opus.
Of course, if you think that this approach is as fast and effective as "vibe coding" as in outsourcing more thinking to the AI, it is not surprising you would conclude the cheaper models were nearly as useful.
I don't know if you are right or not, a lot depends on the constraints of the project and team.
Some of the newer arguably now viable use cases, such as porting a large codebase to Rust, are certainly not going to be as fast with a more manual approach.
Eventually it plateaued and now you can do a decent chunk of your computing on something from 2012.
People keep saying scaling will top out, for example. But scaling keeps stubbornly refusing. New techniques keep coming along too. It's really still exploding into existence and every new generation brings new capability. Eventually it'll clear a ceiling for your key use cases and you'll stop worrying about new models.
It always pays to look back at history and see if you can pattern match.
I sometimes chuck a few tokens to gpt 5.5 and opus 4.8 and they can sometimes solve a problem one of the other models couldn’t, but they’re not like 10x better or anything in my experience. More like 1.2x better
I would not be surprised if it is not as intelligent as the Mythos class models.
I have seen rumors that GPT 6 may release before September. The same person also claimed that a Fable 5.1 checkpoint has been completed a few weeks ago.
How does this go through so many layers of management at a trillion dollar company without who has a say raising this? I simply can't believe how stupid the naming scheme from OpenAI was and continues to be even after they acknowledged it earlier.
Whereas Sol/Luna/Terra reads more like "GPT for hard/medium/basic problems".
Possibilities are endless:
Sounds nice, looks cool. Why not?Past Intel had way better naming.
It'd just be OpenAI GPT 5.6++++++++++++++
OpenAI GptForce 5600 XT
What surprises me is not this, but that OpenAI changed things up without syncing with a GPT 6.
On a more serious note, I can vividly imagine how difficult it would be to find a set of words that could plausibly suggest a relational meaning while remaining non-diminutive in every single model name. Adding to the complexity, it is going to be used globally, and the main competitor already has a successful and fabulous naming scheme.
It sounds like a PR minefield.
It sounded nicer than something like "Luxury" and "Basic".
That’s interesting. My understanding is that:
Sol = Sun Terra = Earth Luna = Moon
So it’s a bit surprising that in Toyota’s nomenclature, Terra is the basic trim instead of Luna.
(Well given the limited amount of things we can deduce from a name)
mini > micro (see, e.g., -skirts, -computers)
micro > nano (see, SI)
so, mini > nano.
I think if you are not seeing reasonable performance in your agent loops as of 5.5, it's likely there is a deficit with how the loop, prompt or tools interact with the environment.
My quota is about to reset. Really can’t wait to use it.
Though its been just 3 days I started using.
Half way through the chat, GPT 5.6 Sol stops and does a safety verification, pretty annoying
Codex CLI just seems faster at coding than Claude Code but Fable is just a level above intelligence wise, it's truly like taking to very very very smart human.
With GPT 5.6 though will be interesting to see if things flip, to have Codex speed (or faster) with Fable level intelligence is a game changer.
In an ideal world we would upgrade 5.4 to 5.6 terra and 5.4 mini to 5.4 luna. But does somebody already have some measurements at least in terms of speed?