MiniMax M2.5: Runtime and Cost per Task Against Claude Opus 4.6

What MiniMax M2.5 is

MiniMax M2.5 is an open-weight agentic model with a 200K token context window — recorded here as 204.8K, the same figure written out — running at roughly $0.30 per million input tokens and $1.20 output through third-party providers. MiniMax describe it as extensively trained with reinforcement learning across hundreds of thousands of complex real-world environments, and claim state of the art in coding, agentic tool use and search, office work and other economically valuable tasks. Those are vendor claims.

MiniMax’s runtime and cost comparison

MiniMax publish a runtime and cost comparison against Claude Opus 4.6. On SWE-Bench Verified, M2.5 consumed an average of 3.52 million tokens per task against its predecessor M2.1's 3.72 million. End-to-end runtime fell from 31.3 minutes to 22.8 minutes — a 37% speed improvement, which MiniMax attribute largely to better parallel tool calling rather than to raw model speed. Claude Opus 4.6 completes the same evaluation in 22.9 minutes, so M2.5 matches it, and MiniMax state the total cost per task is 10% of Opus 4.6's.

The accuracy figures

MiniMax report 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp with context management enabled. The BrowseComp caveat is the vendor's own: context management is a system around the model, so that score describes a configured setup rather than a bare model call. All three are MiniMax's measurements on MiniMax's harness.

LMArena

On LMArena's text leaderboard MiniMax M2.5 has 1391 Elo (rank 179); by category, 1446 Elo in Coding (rank 164), 1393 Elo in Math (rank 174), 1425 Elo in Expert (rank 155) and 1356 Elo in Creative Writing (rank 176).

Two speeds and MiniMax’s pricing

MiniMax release two versions with the same capability. M2.5-Lightning runs at a steady 100 tokens per second and costs $0.30 per million input tokens and $2.4 output; M2.5 runs at 50 tokens per second at half that price. Both support caching. MiniMax state that running M2.5 continuously for an hour costs $1 at 100 tokens per second and $0.30 at 50. M2, M2.1 and M2.5 were released over three and a half months starting in late October. MiniMax also upgraded its VIBE benchmark to a Pro version for this release, and state that M2.5 is served natively at 100 tokens per second.

Training and harness results

MiniMax report training on more than 10 programming languages, including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart and Ruby, across more than 200,000 real-world environments, using an in-house agent-native reinforcement-learning framework called Forge. On SWE-Bench Verified run in other harnesses, MiniMax report 79.7 on Droid against Opus 4.6's 78.9, and 76.1 on OpenCode against 75.9. On BrowseComp, Wide Search and MiniMax's RISE search evaluation it used about 20% fewer search rounds than M2.1. For office work MiniMax built an internal evaluation, GDPval-MM, on which it reports an average win rate of 59.0% against other models; the work covered Word, PowerPoint and Excel financial modelling with input from finance, law and social-science professionals.

Official Resources