One developer beat many chatbots on a reasoning test for 67 cents.
Independent developer Mithil Vakde trained a small model from scratch on the ARC-AGI-1 puzzle set. He reports 44 percent on the public evaluation after 90 minutes on a single gaming graphics card. The code is open, and he puts the whole run at about 67 cents of compute.
Why it mattersReasoning gains have looked like something only enormous budgets can buy. A cheap open result suggests the method may matter more than the money.
Read in the app — free, in 9 languages
Related stories
OpenAI's next model thinks in loops that leave fewer traces.2026-09-03Perplexity cites 215,128 buying guides made by three linked websites.
2026-09-02Anthropic found faults in over 10 percent of its own training tasks.
2026-09-02The best AI agent finished 30% of real lab work
2026-08-29Claude beat human researchers at closing alignment gaps
2026-08-29