Man

@dschwarz26 Grade it as correct. I think enough time has passed. Fable is more intelligent.

GPT-5.6-Sol has been out a week. From Zvi: > In terms of raw intelligence and ‘big model smell,’ and ability to do the hardest things that are intelligence-loaded, Fable still looks like it has a substantial edge... I still consider Fable ‘the best’ model Anyone want to defend GPT-5.6-Sol against Fable, or can I grade this prediction as correct?

@littmath Why does fable do better on frontiermath?

Overall I think 5.6 Sol Pro/Ultra etc. seems to be a substantial step up from 5.5 for math. That said, common interaction pattern is: I ask a question. It thinks for ~100+ minutes and returns a largely inscrutable response. I ask it to explain. It thinks for 20 minutes and says:

@Sigma_BB_Girl @sumukx Why do you hate ant?

Working on a very physics-heavy ML repo, and your post so accuately and eloquently describes my feeling towards 5.6. In this repo, it 100% feels like a regression, doing what feels like prep work for future work, but never doing the actual work. FWIW, the only other OAI model that felt like a regression to me was 5.3-codex. Went right back to 5.2. I absolutely hate Ant, but Fable did not have these issues. Looks likely I will go back to 5.5 or try Grok 4.5

Some thoughts about GPT-5.6-Sol after ~30B tokens: Sol is the most OCD model I’ve used thus far. It very frequently gets one-shotted by random nits in the codebase and writes a bunch of tests to fix it. Even with fast mode, it’s incredibly slow to do this kind of iterative development, especially when builds take really long. This by itself is not a bad thing, but the worst part is that after 2 compactions, it’s chasing the nitpick / useless goals I never told it to accomplish, rather than the main task. This behavior is so bad, I thought I was messing something up and tried codex, pi, and opencode to figure out if it’s a harness issue, but there is no meaningful difference between the three, which leads me to believe this is a model problem. AI code has this weird delayed release effect. You’ll only notice slop code 2 dev cycles into a codebase when you spend more time fighting with the code and on refactors than on shipping features. It’s possible that sol is better than 5.5 a couple cycles in, but tbd. My file deletion experience has also been similar to others: this is a dangerous model to let loose without guardrails. For instance, when performing a routine container upgrade, it accidentally printed out an env secret, then panicked and rotated ALL secrets (this is internal so not public facing, which was also documented), and proceeded to break everything, spending an extra hour fixing everything and redeploying everything else to use the new secrets. It also gets rid of files it doesn’t like. I have no idea why this is, but I think something about the reward model rewarded bookkeeping. Writing is another problem. 5.6 has a huge context bleed effect. It does not know how to write documentation and starts putting the specs in the documentation. If I ask it to develop a user sandbox for isolation, and also ask it to write documentation, it starts talking about specs and sandboxes in user-facing docs, which makes no sense. Fable is somehow much, much smarter in this regard. Frontend design has also not gotten better. Fable is still one generation ahead here. Overall, as a huge 5.5 user, I am not convinced that sol is a meaningful upgrade. It’s possible my practices need to change, but unfortunately it feels like I’m spending longer fighting with 5.6 than I did with 5.5. It’s like the model is so SO smart, but so hard to work with, compared to fable and even grok4.5 surprisingly. It’s clearly intelligent, but also just doesn’t care about what I ask it to do? (Is this supposed to be AGI feels like?) I hope the codex team fixes what possibly is a bad harness setup, because the benchmark numbers show a very different story from what I’m seeing while using the model.

@robbensinger Why is it bad? They are an AI company with safety is one of its core missions.

... Am I weird for thinking this is a remarkably bad sort of thing for Anthropic to release? It looks and sounds like a Claude ad. It is seriously bad to send the message that talk of AI risk, macroeconomic effects, etc. is part of a marketing campaign to boost Anthropic's image.

@SebastienBubeck Why does it do worse than fable on frontiermath?

I think 5.6 might be good at math

@zarazhangrui Everyone should be using Claude tag.

The 3 levels of AI adoption for organizations Most companies are at level 2

@IbrahimDagher20 Seems like Anthropic is the most focused on RSI. They have the biggest and best model in the world even though it's expensive. And from most accounts it has the best intuition and feel of a researcher.

You should think of the current LLM era as an on-ramp to AI, not as something permanent. LLMs will *bootstrap* us there. This is why it’s less important to focus on which models are currently cheaper or more efficient, and much more important to focus on who is going to reach the bootstrapping point first. For example: when we first invented fire (wood-based), that wasn’t hot or efficient enough to smelt iron. But it would be a mistake to try to invent furnaces or engines from first-principles, or to focus on making cheaper, less-good wood fires to capture the fire market. Instead, you use wood fire as a bootstrap: once wood fire is hot enough to make charcoal, you now have something which burns even hotter and is enough to get you to burn bronze, which allows you to make tools that then create blooms, which ultimately lights the fuse of an Industrial Revolution. Similarly, dont focus on which models happen to be more efficient (eg open source). What matters is which company is closest to LLMs that can actually run machine learning experiments near-autonomously, at which point it will be much easier to explore the space of AI at scale and find AI paradigms that learn on the job, are cheaper to train, can be customized to you, etc. The company that hits the bootstrap first will be the company that can churn out new AI paradigms that capture the market, among other things.

@deskmerchant @drydenwtbrown Opus 4.5 was the big change no one saw coming and now with Claude tag it seems clear what the labs want.

any theory as to why these people seem to just be getting privy to these dynamics. why wouldnt satya have known all of this 3 or 4 years ago and been moving in coordination with that worldview since then. why does this seem like recent revelations for everyone that specifically is in a position to be future forecasting ?

TFW you’re legitimately concerned OpenAI and Anthropic will eat the entire economy

@schrockn In terms of capabilities whats ways is it better and worse than fable?

I've seen enough: After a weekend of working with Sol and Fable it's pretty clear to me that OpenAI has taken the lead on coding models. Congrats to the team. You cooked. Looking forward to Anthropic's next move!

@verrsane @tonygentilcore Already happening with claude tag inside slack. They are several months ahead of their competition. Coming to Microsoft teams as well apparently.

the future of AI is non technical folks just opening up PRs instead of waiting for engineers after complaining in slack

@uyintans @ziv_ravid It's very obvious they are all in on RSI/AGI. My guess is they think working on these smaller models is taking away from that mission.

I don’t think it was a mistake, but it was a mistake to focus all of their energy on something so expensive while letting Opus and Sonnet rot. Everyone at Anthropic seems to just use Mythos all of the time without any clue how much it costs outside of there They need this price war

I wonder whether, in retrospect, Anthropic thinks it made a mistake in developing such a large model. It made more sense in February, when there was no model near its performance level, but now, their only option is to keep subsidizing it through the paid plan

@sdmat123 @zugzwangg_ @then_there_was Lol am i talking to a bot? God dammit.

@MsShannon112 @zugzwangg_ @then_there_was This is the same Dario who is fond of proclaiming AI will replace all developers in 6-12 months

I think Anthropic’s hesitation to extend Fable indefinitely on paid plans largely comes down to limited compute. Their priority is large-scale training runs, and they don’t want customer inference to slow their progress towards AGI.

@sdmat123 @zugzwangg_ @then_there_was You watched the Video and still get his point? Being too early can make you go bankrupt. How do you not understand this?

@zugzwangg_ @then_there_was Entirely Dario's fault for not making compute commitments. He explained his risk aversion rationale for this on the Dwarkesh podcast. It's odd - you would think the guy repeatedly proclaiming that AI will eat the economy would not get cold feet about compute.

@then_there_was @AlanRBlair This is someone's WRONG interpretation of Anthropic posted like it was true. This argument could be made for most AI companies BTW.

@AlanRBlair Yes. And control over it.

@DaveRBanerjee Yeh, people don't code. Even claude COwork is quite niche. That's why Fable being 5% better than the next model will not make a difference it's the applications that matter.

I feel like it’s been much harder for me to feel the exponential lately because I haven’t been programming in much In fact, most of my non-programmer friends feel that progress has stalled

@Presidentlin Anthropic is smart they know most of the money will come from non technical people doing technical work with Agents. Claude Cowork, Legal, Design etc is for non technical people. OpenAI has a lot to catchup on.

People don't understand OAI has to get non technical people into this stuff. Burn Codex if you must. The worst customer is the techy Twitter person. Loads up on $200 plans. Maxes them out. Runs on ultra for some reason. No matter what you believe about inference being high margin. The people who fully use the $200 eat into that. The $200 user is also a bad customer because they wouldn't buy $200 worth in API spend. Yet people dare to ask for a $500 plan. Worst customers for the labs. In order - Free users who use this stuff - $20, $200 maxxers - Enterprise Customers who baulk at the standard api After that best are People who pay for this stuff but use it a little. Anthropic is way worse. Fable is going to be a once per week model. For emergencies. You don't want your model to be an emergency model, the company doesn't get hooked on it, and they start thinking they can live without it.

@j_g_allen I disagree. It is more complementary, these products allow people who would never look at Code/Legal/Design etc. to participate.

Been saying this for awhile - that the base models are eating any business they want - and now, finally, the big players w a lot to lose are seeing it and saying something. This is unsustainable for the economy, and it’s impossible to compete bc these AI companies have access to AI frontier models no one else has, months or years before others have access…

@JasonDClinton This was posted 63 weeks ago which suggest Claude Tag was being worked on over a year ago. Is this true?

@tikgiau WHen will you test Fable 5 and 5.6 Sol?

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog edge-bench.org Paper edge-bench.org/paper.pdf GitHub github.com/ByteDance-Seed… Dataset huggingface.co/datasets/ByteD… Details below 👇🧵

@dylan522p If he didn't lie so much anthropic would still be part of OpenAI. Lied to Paul grapham, lied to ilya, lied to dario ,lied to the board. Still probably lies to him employees.

Or he failed to keep it together (Anthropic, Thinky, Core Automation, etc) and bloviated so much about his tech that everyone saw it as existential and it created completion from all the hyperscalers. If he played his cards right I'd bet there'd be no Anthropic or Meta TBD or XAI