Atomic.chat says Qwen 3.7-max beat Opus 4.7 and GPT-5.5 in agentic Tetris-bot test
In a 10-iteration code-and-rewrite challenge, the team let each model build and train a Tetris bot, then compared the final agents.
By Ryan Merket · Published
Why it matters
Operator-style, agentic evals are closer to how teams will actually use frontier models. If Qwen is edging rivals in code-and-iterate loops, stack choices and costs could shift quickly.
Atomic.chat said in a thread on X that Qwen 3.7-max outperformed Opus 4.7 and GPT-5.5 in a real agentic task: write a Tetris bot that plays the game and trains itself.
https://x.com/atomic_chat_hq/status/2057581603811901882
Atomic.chat described a closed-loop setup where each model could read its own code, run benchmarks, and rewrite itself across multiple rounds. In a reply to Thomas Truszkowski | AI Geek (sarcasm included) (@Thomas_AI_geek), the account clarified it ran 10 iterations per model before comparing the resulting bots.
The company did not publish detailed metrics in the thread but shared a video of the bots and said the run was costly, responding to erik@try.works (@trydotworks) that it was "expensive as a hell." Atomic.chat also linked to its site for more context: atomic.chat.