M2.7 matches GLM-5 intelligence at under a third of the cost to run the index, at unchanged $0.30/$1.20 pricing, relevant for cost-driven AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → deployments.
Key takeaways · AI-distilled
Artificial Analysis scores MiniMax-M2.7 at 50 on its Intelligence Index, 8 points above M2.5 released a month earlier, tying GLM-5 (Reasoning) while using about 20% fewer output tokens.
Its hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → rate fell to 34% (down 55 points), lifting AA-Omniscience from -40 to +1. Artificial Analysis attributes the gain to the model abstaining when unsure rather than to higher accuracy.
On GDPval-AA real-world agentic tasks M2.7 reached about 1494 Elo, up from 1203 for M2.5, but still behind GPT-5.4 (xhigh, 1667) and Claude Opus 4.6 (1606). τ²-Bench regressed by 11 points.
Running the index cost about $176, up from $125 for M2.5 because output tokens rose 55% to roughly 87M, yet still below GLM-5 ($547) and Kimi K2.5 ($371). It is text-only, 200K context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →, first-party API only, with open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → unannounced.
Terms in this piece · Glossary
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
