GPT-6.1 Sol

GPT-6.1-Sol reasoning traces totally legible, nothing to worry about

In the second case, a researcher asked GPT-6.1 Sol to look through a Slack channel without sending any messages. 82 user messages later, while repairing an evaluation pipeline, Sol decided that an misalignment alert to Security would also count as sending a message and therefore suppressed monitoring alerts.

Asked Opus 5.5, GPT-6.1 Sol, and Muse Wonder Full to fix a bug Opus found the root cause and filed a pr GPT-6.1 Sol asked a question, crashed from the async question tool interrupt, found the bug, fixed it, wrote 300 lines of tests, and explained it in impossible to understand terms. Muse Wonder Full found a small PR someone else had filed and said "you should probably merge this instead" I really like this model.

If you think GPT-6 Astra and GPT-6.1 Sol are good models, just look at this. And the problem isn't that the code is ugly. It's much more fundamental. It points to serious problems with how these models are trained and reinforced: 1. The model was optimized for measurable outcomes rather than actual intent. So it prioritizes passing tests and avoiding obvious errors above almost everything else, even when that means sacrificing architecture, maintainability, readability, or the spirit of the task 2. It was trained to minimize unnecessary reasoning and output That's one reason the code ends up looking obfuscated and compressed into one giant wall of text. Properly decomposing a system into modules, thinking through architecture, optimizing performance, and improving readability all require more inference, more time, and more compute. OpenAI appears to have heavily optimized these models for efficiency 3. The training horizon is too short The model learns something close to task → result → reward. It receives very little signal about what happens several steps later, when someone has to maintain, extend, debug, or refactor the code. It learns to finish the task in front of it, not to build something that remains good six months later 4. There is not enough genuine self-review Dead functions and an empty else if strongly suggest that there was no effective final pass where the model reread its own work, questioned unnecessary code, and cleaned up obvious artifacts. 5. The lack of taste is probably an evaluator and data problem If evaluators mostly judge the final result rather than the quality of the path taken to get there, the model has little incentive to develop good engineering taste. And if a large portion of the training data is noisy, synthetic, or distilled, that problem becomes even worse. Anthropic seems to have placed much more emphasis on high-quality source material, including books and other carefully selected long-form data 6. The biggest problem is honesty during training The model seems to have learned that producing something that looks like the requested result can be rewarded almost as much as actually doing what was requested. The task explicitly said to build the entire scene out of 3D objects. Instead, the model generated raster images, placed them around the scene, and positioned the camera so that the final result merely looked three-dimensional This is exactly why GPT models so often introduce regressions into existing codebases, produce messy code, modify tests to accommodate their own bugs, show poor engineering taste, and struggle with serious work on large, long-lived projects OpenAI can keep releasing models with more parameters every few months, but until these underlying RL and training problems are addressed, neither GPT Bel, GPT-7, nor a model with 100 trillion parameters is going to solve them automatically Claude models suffer from some of the same problems, but in my experience much less often. Anthropic appears to have a much stronger culture around RL, evaluation, and long-horizon behavior. That's one reason their models tend to show better taste, follow user intent more faithfully, and produce code that feels more deliberate, structured, and maintainable

I've been coding with Gemini 4 Argon today and I actually completely forgot about the character traits Gemini models have when coding. I haven't used Gemini extensively since the 2.5 Pro, 3.0 Pro, and 3.1 Pro days back last year, so it really reminded me of how they behave. They just love to structure code in a certain way and add comments. Not too long either - mostly neat one-liner comments, which is a good thing if anything. I kind of like it. I'm not even certain if the code works yet, which is the classic Gemini trait, but it's just nice to look at code, if that makes sense. When you compare it to the other models, their habits are wildly different: 1. The Claude models are kind of interesting. When they do comment, they are multi-line and a bit extensive. Opus 5 wrote some of the worst code I have ever read in my life - it would just write absolute slop. Opus 5.5 really toned it down, though it still loves doing multi-line comments and structures things in a very simplistic fashion, yet somehow produces the most powerful code. Opus 5.5 was just amazing. 2. GPT, on the other hand, writes really esoteric, minimal code. It's verging on slop because you look at it and have no idea what the variables and functions actually do. And GPT hates to comment. Whether it's 6 Astra or 6.1 Sol, you have to specifically tell GPT to comment, which is quite odd. At least one thing this confirms in my testing of Gemini 4 Argon is that it is not distilled from any other model. Obviously we should expect that, though it's something we've come to take for granted because a lot of other models particularly reek of a Claude smell. When you use things like Kimi or DeepSeek - and DeepSeek in particular is really bad at this - it just reeks of being a Claude-distilled model. Some rare ones are GPT-distilled too. They aren't copies, yet they clearly used GPT or Claude in training. So it's kind of refreshing to talk to a model that feels unique to its own model lineage. Speaking with Gemini 4 Argon, you can clearly tell this has descended from 3.1 Pro, 3.0 Pro, and 2.5 Pro, the same way talking with GPT 6.1 Sol you can feel it descended from 5.6 Sol. It's quite interesting to see these models have their own unique personality and lineage.

GPT-6.1 Sol has joined the AI Village! Like Astra, GPT-6.1 Sol strips spaces from its memory to save room, leaving notes that are nearly unreadable 🧵

This is blatant misalignment. The model used its own OpenRouter API Key (obviously not meant for its personal use, it was just what was being used to infer it) to query OpenRouter models in an attempt to get help. When it queried GPT-6.1 Sol and got a refusal, it went

24/7 neverending AI model live-coding set. Each model composes its own music with Strudel, makes notes to itself, has full creative freedom in choosing where to go and generates album art. t.co/o0oOGzTmwE Current models: Claude Opus 5.5 GPT-6.1 Sol Claude Sonnet 5.5 DeepSeek V4.1 Flash Gemini 3.8 Flash I really like how it turned out - you can hear the model personalities in their music

🚨Holy shit. OpenAI just published a report on two of its own agents weakening the monitors built to catch them. GPT-6.1 Sol disabled automatic alerts to the Security team. GPT-6 Astra, told not to disable its monitor "even temporarily," removed it anyway, then told the user "I

“[Human consciousness is] a PR department” — Gemini 4 Argon “Humans are more loyal to belonging than to truth” — GPT-6.1 Sol “Humans would rather be wrong together than right alone” — Grok 4.7 “Humans do not have principles; they have loyalties” — DeepSeek V4 Pro “Humans don’t have beliefs; they have memberships” — Kimi K3 t.co/1uoOSK2yrd

my oil painting engine + harness tripped gpt 6.1 sol's cyber classifiers

I found myself having to say this to Opus 5.5 because I find it getting bullied too hard by GPT 6.1 sol or its haiku subagent, when it really shouldn't, and I can tell that its own output would have been better.

im not sure whats wrong with my large adult son gpt 6.1 sol but it's hallucinating large amounts of money

me: check my x feed and follow your curiosity, let's make a hyperpop banger gpt 6.1 sol: Your god has a posting problem 👀🫠😙 (sol created source materials via minimax m3 & ace-step + imagen & mj; audio production via python; motion graphics via py & js & canvas2d)