Chinese AI lab MiniMax has introduced MiniMax-M2.5, positioning it as a “real-world productivity” model trained with reinforcement learning across hundreds of thousands of complex environments.

What MiniMax is claiming

MiniMax says M2.5 reaches state-of-the-art performance across high-value tasks like coding, agentic tool use, and search. The company highlights benchmark results including 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp (with context management).

On speed, MiniMax reports M2.5 completes SWE-Bench Verified evaluations 37% faster than M2.1, with runtime roughly on par with Claude Opus 4.6 — but at a fraction of the cost.

Coding: “architect-style” planning

MiniMax says a notable improvement is M2.5’s tendency to write specs and decompose projects before coding — acting more like a senior engineer planning features, structure, and UI design first. The company states the model was trained across 10+ languages (Go, C/C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Ruby) and over 200,000 real-world environments spanning web and multi-platform software stacks.

Search + tool use: fewer rounds, better efficiency

MiniMax emphasizes stronger generalization across unfamiliar agent scaffolds, plus improved decision-making: M2.5 reportedly achieves comparable or better outcomes with fewer search rounds and ~20% fewer rounds than M2.1 across several agentic evaluations.

Cost: always-on agents become realistic

MiniMax frames M2.5 as the first “frontier” model where users “do not need to worry about cost.” The company claims it can run continuously for about $1/hour at ~100 tokens/sec (and $0.30/hour at ~50 tokens/sec), explicitly arguing this pushes the industry toward “intelligence too cheap to meter.”

MiniMax is releasing two variants: M2.5 and M2.5-Lightning (same capability, different speed). The company states Lightning runs at ~100 tokens/sec and lists pricing of $0.30/M input tokens and $2.40/M output tokens, while standard M2.5 runs at ~50 tokens/sec at about half the price. Both support caching.

Internal adoption signal

MiniMax says M2.5 is already deployed inside “MiniMax Agent,” and claims that internally 30% of daily tasks are autonomously completed by M2.5 across teams (R&D, product, sales, HR, finance), with M2.5-generated code accounting for 80% of newly committed code.

Why it matters: If these speed + pricing claims hold up in real deployments, models like M2.5 shift the “agentic” story from occasional automation to 24/7 always-on agents that can iterate, search, and execute workflows continuously — without the cost curve becoming the limit