# MiniMax M3 on the Cotera agent benchmark > Thinks in `` blocks. Sometimes forgets to write the answer afterward. Source: https://cotera.co/benchmarks/models/minimax-m3 --- MiniMax's M3 — the November 2026 reasoning model out of the Chinese MiniMax lab. Tested via the first-party API. **Score:** 3/5 benchmarks passed · **Cost across the matrix:** $1.10 ## Strengths - Sales pass was thorough: 20 tool calls verifying Crunchbase data against multiple sources, $0.254. Most agent-loop-style behavior of any model that succeeded here. - Web Scraping ($0.044) and Marketing ($0.200) passed cleanly with no envelope issues. ## Watch-outs - M3's signature failure mode: opens a `...` block, plans the answer, then closes the run before writing the answer. Both CX and Coding failed this way. The thinking is visible; the output isn't. - This is a known issue with M3's chat-template — the model treats `` as a non-streaming scratchpad and sometimes doesn't transition out of it before hitting stop tokens. - Coding cost $0.508 to produce no output. M3 read the Stripe docs, planned the function in a `` block, then stopped. You're paying for thinking the user never sees.