MiniMax M3.1 Flash is out, and M Plan calls to M2.5 and M2.7 now run on it

MiniMax released M3.1 Flash on 19 October 2026. The Chinese AI company says the model is built for agent work, coding, computer use and security tasks, and it priced it well below most frontier APIs. The launch also changes what some existing MiniMax customers are running, whether or not they asked for it.

If you already use MiniMax

The model list carries a retirement notice for the M2.5 and M2.7 families:

• 19 October 2026: the models “will no longer be available to new users.”
• 18 November 2026: the models “will be retired”, and API requests to them “will no longer work.”

What happens in between depends on how you pay.

• M Plan and Token Plan users: “Starting October 19, 2026, all requests to M2.5 and M2.7 models under these plans will automatically be routed to M3.1 Flash.” Old model names keep working, and MiniMax says plan allowances stay the same. In practice that means a request that names M2.7 is now answered by a different model. If you have prompts, parsers or evals tuned to M2.7 output, rerun them.
• Pay-as-you-go API users: existing users can keep calling M2.5 and M2.7 until 18 November, then the calls fail. Plan the switch before then.

One oddity: the same docs page already lists MiniMax-M2.7, M2.5 and their highspeed variants in its “Retired Models” table. Go by the dated notice above, and test your own calls rather than assuming either way.

What M3.1 Flash is

According to MiniMax’s text-generation guide, MiniMax-M3.1-Flash has a 1,000,000-token context window and accepts text, image and video input. It always thinks: an effort setting takes low, medium, high, xhigh or max, and “when omitted, the default is max.” If you are moving latency-sensitive calls over, set effort explicitly rather than inheriting the most expensive level.

MiniMax also says “Space Bunny”, an anonymous model that it says processed more than 200 trillion tokens, “was an early version of MiniMax M3.1 Flash.”

Price

The pricing page and the launch post agree on a flat rate per million tokens:

	Input	Output	Cached input
MiniMax-M3.1-Flash	$0.15	$0.60	$0.015

For comparison, the same page lists MiniMax-M3 at $0.30 input and $1.20 output per million tokens for prompts up to 512k tokens. MiniMax claims that for agentic tasks, M3.1 Flash’s “overall usage cost is 70% lower than its predecessor”, and that output speed “can reach 200+ TPS” with median speeds of 170+ tokens per second. Those are vendor figures.

Benchmarks, all from MiniMax

The launch post reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, 76.2 on BankerToolBench and 82% on CyberGym. It says the model “ranks first among Chinese models” on OSWorld V2 and Agents’ Last Exam. We have not run any of these ourselves.

The one outside number so far comes from Artificial Analysis, which gives the model 42 on its Intelligence Index and says it cost $0.22 per task to evaluate on that index. Artificial Analysis also says the model generated 180M tokens while running the index, which it calls “very verbose” next to a median of 99M. That matters for real bills: a low per-token price can still add up if the model writes a lot.

Open weights: not yet

The launch post says “Model weights will be released later,” with no date. The text-generation guide already describes M3.1 Flash as “an ultra-efficient open-weight model”, but Artificial Analysis lists it as proprietary and says the weights are not publicly available. Until a download exists, treat it as an API-only model. MiniMax also says a larger “Pro” model “will also arrive later.”

Where to get it

MiniMax says M3.1 Flash is available now through the MiniMax API, its M Plan subscription and its MiniMax Code app. The text-generation guide shows calls through MiniMax’s own endpoints.

For the wider context on retirements forcing migrations, see our piece on why your model has a retirement date, and this week’s notice that OpenAI will shut down GPT-4o and GPT-4.1 API snapshots in January 2027.