A Chinese video-generation startup just quietly beat Claude Opus at coding

While the AI world kept watching OpenAI and Anthropic, a Chinese lab quietly shipped a model that meets the high bar for automated software engineering. MiniMax, known until recently for its video generation models, released M2.7, a language model that scores 56.22% on the SWE-Pro benchmark. The company says that result is close to Claude's Opus family. MiniMax launches M2.7 model with strong software…

SWE-Pro tests how well a model handles real-world software engineering tasks: fixing bugs, writing patches, navigating codebases. Only a handful of models have cracked 50% on this benchmark, and most of those belong to labs with much larger budgets. MiniMax, a startup founded by former top researchers from Huawei and Microsoft, is now playing in that league. Cognition's new coding agent scores near frontier…

But M2.7 is not just about code. The benchmark sheet is dense with claims that cut across domains: end-to-end project delivery on VIBE-Pro (55.6%), deep system understanding on Terminal Bench 2 (57.0%), and a top score on office productivity. On GDPval-AA, which measures document editing proficiency across Excel, PowerPoint, and Word, M2.7 achieved an Elo of 1495, the highest among open models, according to MiniMax.

Skill adherence at scale

One of the more revealing numbers in MiniMax's marketing deck is a 97% skill adherence rate on what the company calls 'complex skills', tasks requiring more than 2,000 tokens of instructions. In a world where LLMs are increasingly tasked with multi-step, tool-augmented workflows, the ability to follow long, intricate instructions without hallucination or drift may be more valuable than a high score on a static math benchmark. Your AI agent passed the test by accident. Now there's…

MiniMax also highlighted performance on the MMClaw benchmark, which evaluates models within the OpenClaw agentic framework, an increasingly popular sandbox for testing how models interact with software tools. On that benchmark, M2.7 approaches the performance of Sonnet 4.6, one of Anthropic's latest models.

The company is also pushing a narrative around self-evolution and multi-agent collaboration, two buzzwords that, in practice, refer to the model's ability to architect its own Agent Harness and coordinate sub-tasks across multiple inference calls. MiniMax claims M2.7 can autonomously build complex Agent Harnesses to complete end-to-end productivity tasks. If true, this would reduce the need for engineers to hand-craft scaffolding for every new automation pipeline. These researchers found a way to make AI agents think…

Two tiers, same result

MiniMax is releasing M2.7 in two inference tiers: the standard model and a highspeed variant dubbed M2.7-highspeed. The company promises identical output quality on both, with the Highspeed edition trading latency for throughput. Both tiers are accessible via the MiniMax API and support automatic caching, a feature that can dramatically cut costs for repeated prompts.

The model is also available through MiniMax's own agent platform, which eliminates the need for coding to experiment with the model. For teams already using AI coding tools like Cursor or Windsurf, MiniMax is offering direct integration.

The pricing remains unchanged from earlier MiniMax models, which positions M2.7 as a potential option for teams that want near-state-of-the-art code generation without the per-token cost of the top-tier Western labs. MiniMax's new M2.5 coding model tops the benchmark at…

A quieter kind of race

M2.7's release comes at a time when the LLM race has split in two directions. On one side, labs like OpenAI and Anthropic compete on general intelligence, multimodal capability, and reasoning chain depth. On the other side, a growing number of challengers, DeepSeek, Qwen, Mistral, and now MiniMax, focus on targeted, high-value slices of the market where a slightly lower score on a broad benchmark is offset by superior performance on specific tasks, lower cost, or better tooling integration.

The explicit comparison to Claude Opus and Sonnet is a tactical choice. By naming names and citing benchmarks, MiniMax is telling the market: you don't need to pay for the most expensive frontier model to get near-frontier performance on the tasks that matter for your engineering team. China's MiniMax just open-sourced a 1M-token model that…

There is, as always, a fine print question. MiniMax has not released the full evaluation methodology for its internal benchmarks, and independent verification will take time. But the numbers, if replicable, represent a significant step for an open model from a lab that, until this week, was primarily associated with generative video rather than code generation.

Office work and identity preservation

Beyond engineering, M2.7's strong performance on office productivity is noteworthy. The Elo score of 1495 on GDPval-AA positions it ahead of many general-purpose models on document editing, a domain that, while less glamorous than code generation, represents a massive addressable market. Enterprises that rely heavily on automated report generation, spreadsheet manipulation, and presentation design may find M2.7 particularly useful.

MiniMax also touts the model's 'identity preservation' capabilities, a claim that the model maintains consistent character and tone across extended conversations, which the company says opens up possibilities for interactive entertainment and customer-facing chatbots that require long-term coherence. Your AI assistant forgets you every morning. This…

The package, taken together, is a model that MiniMax clearly intends to serve as a general-purpose agent base, not just a code generation tool. The combination of high code benchmark scores, office productivity, and agentic capability with near-perfect instruction adherence suggests the lab sees its competitive advantage not in raw intelligence but in reliability and breadth of tool use.