MiniMax drops M2.5 with frontier speed at a fraction of the cost

MiniMax just rolled out MiniMax-M2.5, a frontier-class model explicitly tuned for real-world agentic workflows, software development, and everyday office productivity.

While the industry has spent months debating benchmark marginalia, MiniMax made a bold play on economics and raw speed: running M2.5 continuously at 100 tokens per second costs just $1 an hour. It also put up formidable numbers, hitting 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp with context management.

Architect-grade coding and trimmed reasoning steps

M2.5 was pre-trained and reinforced across more than 200,000 real-world environments and over 10 programming languages, from Go and Rust to TypeScript and Python.

MiniMax noted an interesting emergent behavior during training: before touching code, M2.5 instinctively drafts specifications like a seasoned software architect, mapping out architecture, features, and UI constraints first. On actual coding agent scaffolds like Droid and OpenCode, M2.5 scored 79.7 and 76.1 respectively, edging slightly past Claude Opus 4.6.

More importantly, it works noticeably faster. M2.5 finished the SWE-Bench Verified test suite in an average of 22.8 minutes, matching Opus 4.6 at 22.9 minutes, while burning about 3.52 million tokens per task. That represents a 37% speedup over M2.1.

In search and office tasks, M2.5 showed better token discipline. Across interactive search evaluations like BrowseComp and MiniMax's internal RISE benchmark, it resolved tasks with roughly 20% fewer rounds compared to M2.1. In workplace scenarios covering financial modeling in Excel, presentations, and documents, MiniMax collaborated with domain experts, achieving a 59% win rate against mainstream models on its internal GDPval-MM benchmark.

Cutting API costs by an order of magnitude

MiniMax is offering two flavors of the model: M2.5 and M2.5-Lightning. Both have the same underlying weights and capabilities, but Lightning delivers 100 tokens per second while standard M2.5 runs at 50 TPS.

M2.5-Lightning costs $0.30 per million input tokens and $2.40 per million output tokens. The standard 50 TPS version cuts that price in half. Put plainly: four M2.5 instances running non-stop for a full calendar year would run you around $10,000. MiniMax claims its output costs are between one-tenth and one-twentieth of Opus, Gemini 3 Pro, and GPT-5.

Under the hood, MiniMax attributes this leap to scaling reinforcement learning. It built Forge, an internal agent-native RL framework that decouples training engines from agent runtimes, speeding up sample training roughly 40x through tree-structured sample merging. It paired that with its existing CISPO algorithm for MoE stability and introduced process reward models to properly credit long agent reasoning paths.

M2.5 is already live in the MiniMax Agent product, packaging these abilities into workplace skills.

Why it matters

The real bottleneck preventing autonomous agents from taking over repetitive work has never been pure intelligence; it is unit economics. Firing off millions of tokens per task with top-tier western models rapidly balloons API bills to unsustainable levels.

MiniMax is offering an appealing proposition here: tier-one coding and search performance at a fraction of the price, running at native 100 TPS. For developers building always-on code reviewers, continuous vulnerability scanners, or deep research agents, this fundamentally alters the math. You can now afford to let an agent run loops without watching your credit card limit evaporate.

We think this positioning is remarkably sharp. MiniMax avoided generic chatbot hype and focused squarely on agent efficiency and execution cost. That pricing will put undeniable pressure on frontier labs to lower their agent inference rates. The caveat is that real production environments are dirtier than benchmarks, and teams will need to evaluate how M2.5 handles messy private codebases. But for anyone building token-heavy agents today, it is well worth spinning up an API key.