MiniMax released M2.5, achieving 80.2% on SWE-Bench Verified at 1/20th the cost of Claude Opus 4.6.
The AMW Read
MiniMax's M2.5 demonstrates a significant shift in the capability-to-cost ratio, validating the 'cost-collapse' pattern and pushing the frontier of reasoning-model economics via efficient parameter activation.
NoveltySignificance
Foundation Models · Player MapScaling Laws
MiniMax released M2.5, achieving 80.2% on SWE-Bench Verified at 1/20th the cost of Claude Opus 4.6. Activating only 10B parameters, it provides frontier reasoning for just $0.15 per 1M input tokens. This shifts AI from a chatbot to a 24/7 worker, allowing 4 agents to run for $10,000 annually. As 80% of internal code is now model-generated, high-end intelligence is finally too cheap to meter. 🚀
