In this article
  1. What happened
  2. The prices, per 1M tokens
  3. What MiniMax’s benchmarks show
  4. Who has to act: the M2.5 and M2.7 retirement
  5. Free on Ollama Cloud until October 22
  6. What to do

TL;DR: MiniMax released M3.1 Flash on October 19, 2026, a low-cost model for coding and agent work, at a flat $0.15 per million input tokens and $0.60 per million output tokens. If your app calls MiniMax’s older M2.5 or M2.7 models on pay-as-you-go billing, those calls stop working on November 18, 2026. M Plan and Token Plan requests to those models are already being sent to M3.1 Flash. Test the new model on your own tasks before the cutoff.

What happened

MiniMax announced M3.1 Flash on October 19 in a blog post and on X. It says “Space Bunny,” an anonymous model that processed more than 200 trillion tokens, was an early version of it. M3.1 Flash is available now through the MiniMax API, MiniMax’s M Plan subscription and its MiniMax Code desktop app.

What MiniMax’s API docs list for it:

  • A 1,000,000-token context window.
  • Thinking depth set with an effort parameter: low, medium, high, xhigh or max. If you leave it out, the default is max.
  • Text, image and video input.
  • The model name MiniMax-M3.1-Flash, on an Anthropic-compatible endpoint (https://api.minimax.io/anthropic, which MiniMax recommends) and an OpenAI-compatible one (https://api.minimax.io/v1).

On open weights, the two announcements differ. MiniMax’s blog says the weights “will be released later.” Its X post gives a date: “Model weights will arrive on Oct 23.”

The prices, per 1M tokens

Model Input Output Cached input (read)
MiniMax-M3.1-Flash $0.15 $0.60 $0.015
MiniMax-M3 (up to 512K input tokens) $0.30 $1.20 $0.06

Source: MiniMax, Pricing Overview (checked Oct 20, 2026). These are standard rates. M3 costs twice as much above 512K input tokens, and its Priority tier is 1.5x standard.

Per token, M3.1 Flash costs half of M3’s standard rate. MiniMax also says that on agentic tasks, total usage cost is 70% lower than its predecessor’s, and that in everyday use output speed can reach 200+ tokens per second, with a median above 170. Those are MiniMax’s figures.

For comparison, DeepSeek’s deepseek-flash costs the same $0.15 and $0.60 off-peak, but $0.30 and $1.20 during DeepSeek’s peak hours: 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays (9 a.m. to noon and 2 to 6 p.m. Manila time). DeepSeek’s cached input is cheaper, at $0.003 off-peak and $0.006 peak. We covered that model in DeepSeek’s move to V4.1-Flash-1017.

What MiniMax’s benchmarks show

In MiniMax’s table (the image above), M3.1 Flash scores 73.7 on DeepSWE v1.1, 61.7 on OSWorldV2, 62.2 on Automation-Bench v1.0.6, 76.2 on BankerToolBench and 82.2 on CyberGym. The same table shows DeepSeek V4.1-Flash-1017 ahead on DeepSWE v1.1 (77.0 vs 73.7), Terminal Bench 4.0 (42.3 vs 38.1), NL2Repo-Bench (68.2 vs 61.2) and BabyVision (92.5 vs 91.0), and Opus 5 leading most rows. These are MiniMax’s own results. We haven’t reproduced them.

Who has to act: the M2.5 and M2.7 retirement

MiniMax’s model list sets this schedule for the M2.5 and M2.7 model families:

  • October 19, 2026: no longer available to new users.
  • November 18, 2026: retired. For pay-as-you-go API users, “Once retired, API requests to these models will no longer work.”
  • M Plan and Token Plan: since October 19, requests to M2.5 and M2.7 “will automatically be routed to M3.1 Flash.” Old model names keep working, and plan allowances stay the same.

The same page’s “Retired Models” table already lists M2.5, M2.7 and their highspeed versions as retired. We go by the dated notice above it, which gives existing pay-as-you-go users until November 18.

Free on Ollama Cloud until October 22

Ollama is offering M3.1 Flash free on its cloud for three days. Its model page lists the free period as running until October 22, 2026, 11:59 p.m. UTC (7:59 a.m. October 23, Manila time), with the same $0.15 and $0.60 per million tokens after that. Ollama says the model is hosted in the US with zero data retention. You can run it with ollama run minimax-m3.1-flash:cloud, or start Claude Code on it with ollama launch claude --model minimax-m3.1-flash:cloud. Ollama’s page lists the model at 428B parameters with a 976K-token context window.

What to do

  1. Search for the old model names. If you pay as you go, look through your code, configs and no-code automations for MiniMax-M2.5 and MiniMax-M2.7, including the -highspeed versions. Switch them to MiniMax-M3.1-Flash or MiniMax-M3 before November 18.
  2. Recheck plan users’ outputs. On M Plan or Token Plan, your old model names have quietly run on M3.1 Flash since October 19. Spot-check results from the last day, especially anything that parses the model’s output.
  3. Set effort yourself. The default is max, the deepest thinking setting. For routine jobs like sorting or tagging, test a lower setting and compare quality and token use.
  4. Test on your own tasks. Benchmarks are MiniMax’s own. The free Ollama window, which runs until October 22, 11:59 p.m. UTC, is a cheap way to try it in Claude Code. If you run DeepSeek during its peak hours, M3.1 Flash’s input and output rates are half DeepSeek’s peak price. Off-peak, they cost the same. DeepSeek’s cached input is cheaper either way.
  5. Don’t plan around the weights yet. MiniMax’s X post says October 23, and its blog only says “later.” At 428B parameters, by Ollama’s listing, self-hosting will need serious hardware.

If you’re moving coding work off a plan that keeps hitting its caps, our guide to Claude and Codex usage limits explains how those limits work.

Sources & method

Sources

Method. Last checked October 20, 2026. Prices, dates, model details and quotes come from MiniMax’s blog, X post and API docs, Ollama’s model page and post, and DeepSeek’s pricing page. Benchmark scores, speed and cost-saving claims are MiniMax’s and were not reproduced by us. Manila times are converted from the UTC times in the sources. We haven’t tested the model ourselves. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.