Shanghai AI startup office at night

China’s open-weight labs are now releasing on each other’s heels. On Monday, Hong Kong-listed AI startup MiniMax (0100.HK) launched an updated M3.1 Flash, claiming “generational” improvements in agentic tasks, programming, and cybersecurity — with online inference speed nearly tripled and usage costs cut by 70%.

The launch lands just two days after DeepSeek’s V4.1-Flash-1017, which claimed 77.0% on the DeepSWE coding benchmark at an effective cost of $0.017 per million tokens. The two announcements read like bids in the same auction.

What MiniMax is claiming

Per the company’s launch blog, M3.1 Flash is built for agentic work:

  • Speed: output speeds of 200+ tokens per second in everyday use, with median (P50) speeds consistently above 170 TPS, via more aggressive model sparsity.
  • Security: 82% on CyberGym for vulnerability discovery and validation, plus strong internal-evaluation results on code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering.
  • Agentic benchmarks: tested on Automation-Bench (complex agentic task execution), DeepSWE (real-world coding), and Agents’ Last Exam (computer use on economically valuable tasks).
  • Price: flat $0.15 per million input tokens, $0.60 per million output, $0.015 cached input — a 70% cost reduction versus its predecessor for agentic tasks. (The older M3 listed at $0.30/$1.20.)

The model is available now through the MiniMax API, its M Plan subscription, and MiniMax Code, and is free via Vercel’s AI Gateway until October 20. Model weights “will be released later” — so for now this is open-weight in promise more than in practice.

The price war is the story

Consider the sequence: on September 27, MiniMax quietly slipped the M3.1-Flash-Preview into its MiniMax Code tool with no model card, no benchmark table, and no per-token price at all. Three weeks later, with DeepSeek having just reset expectations at $0.017-per-million-tokens territory, the full launch arrives with flat pricing published on day one.

That’s the pattern to watch. Chinese open-weight labs are no longer competing primarily on benchmark supremacy — they’re competing on tokens per dollar at agent scale. When the model doing your multi-step coding agent’s work costs a fraction of last month’s price, the economics of agent deployments change faster than the leaderboard does.

The caveat stands as always: every benchmark figure here is company-reported, and “generational” is MiniMax’s adjective, not an independent finding. We’ll be watching for third-party runs — and for whether the weights actually ship.

Sources: Reuters (Oct 19, 2026); MiniMax launch blog; Vercel AI Gateway. Benchmark and pricing figures are the company’s own and await independent verification.