Released 19 October 2026
MiniMax M3.1 Flash: what it costs, how it scores, and how to run it
A 1,000,000-token context window, $0.15 per million input tokens, and a measured cost of$0.22 per agentic task. Here is the honest version: which numbers to trust, which configuration steps break, and when a different model is the better call.
Start here
Three ways to run MiniMax M3.1 Flash
Pick the path that matches how you work. All three reach the same model; they differ in billing, control and setup effort.
M Plan subscription
Go, Explore and Build tiers at 1x, 3x and 7.5x relative usage, governed by a 5-hour and a weekly window rather than a monthly bucket. Works inside MiniMax Code and, with a Subscription Key, inside the coding agents you already use.
MiniMax API
Anthropic-compatible and OpenAI-compatible endpoints at api.minimax.io, with full control over reasoning effort, prompt caching and service tier.
Ollama cloud
One command: ollama run minimax-m3.1-flash:cloud. Also launches Claude Code, OpenCode, Hermes and OpenClaw against the model directly.
The wider lineup
M3.1 Flash is one model in a much bigger catalog
MiniMax ships text, video, speech and image models on the same platform, and the M2.x family is being retired. Here is what is live and what is on its way out.
Text and coding
MiniMax-M3.1-Flash is the current flagship: 1M context, always-on reasoning with an adjustable effort level. MiniMax-M3 is still available but doubles in price above 512k input.
Video
MiniMax-H3 does text-to-video, image-to-video, first-and-last-frame and reference-to-video at up to 2K, billed per second of output. H3-Max is the fast variant.
Speech and image
speech-2.8-hd and speech-2.8-turbo for text-to-speech, voice cloning and voice design; image-01 and image-01-live for text-to-image and image-to-image.
On the way out
M2.5 and M2.7 are closed to new users and fully retire for pay-as-you-go users on 18 November 2026. M Plan subscribers are rerouted to M3.1 Flash automatically.
Same measurement run
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash
Both rows come from one Artificial Analysis run on 21 October 2026. Note the caveat on the speed row before you read it as a verdict.
| Metric | M3.1 FlashMiniMax, 19 Oct 2026Recommended | V4.1 FlashDeepSeek, 10 Sep 2026 |
|---|---|---|
| Intelligence Index | 42.71 | 39.46 |
| Cost per task | $0.22 | $0.571 |
| Input / output per 1M | $0.15 / $0.60 | $0.30 / $1.20 |
| Median output speed | 104.7 t/s | 233.5 t/s |
| Context window | 1,000,000 | 1,000,000 |
| Weights | Not released | Open, MIT |
The guide
What this site covers that the parameter cards do not
Cost, in tasks not tokens
Prompt-cache arithmetic, the priority tier, and the fact that reasoning effort defaults to max — the three things that actually decide your bill.
Benchmarks you can calibrate
The official scores, the third-party scores, and the reason a headline benchmark in this release is documented as flawed.
The steps that break
Environment variables that override your settings file, the model suffix that unlocks long context, the Codex version floor, and the retired preview name.
Where it falls short
A measured hallucination rate, no independent reproduction of the speed claims, and no open weights yet.
Real workflows
The four tasks MiniMax documents, with acceptance checklists that separate a working deliverable from a convincing demo.
Sourcing on every page
Official, third-party, ours, or community — labelled, dated and linked. When we could not verify something, we say so.
FAQ
Questions people ask about MiniMax M3.1 Flash
Short answers first. The linked pages have the full detail, the sources and the caveats.
What is MiniMax M3.1 Flash?
A multimodal reasoning model from MiniMax, generally available since 19 October 2026, built for agentic work, coding, computer use and cybersecurity, with a 1M-token input context window.
How much does it cost?
$0.15 per million input tokens, $0.60 output and $0.015 cached input, with no long-context price step. Artificial Analysis measures a weighted cost of $0.22 per agentic task.
Can I use it in Claude Code or Codex?
Yes, and both have official MiniMax guides. Claude Code uses the Anthropic-compatible endpoint; Codex uses the OpenAI-compatible one plus a model catalog file. Both are covered step by step here.
Are the open weights available yet?
No. The launch post says weights come later, the Hugging Face repository returns 404, Artificial Analysis records isOpenWeights as false, and MiniMax self-hosting docs do not cover M3.1.
Is it as good as the frontier models?
It is competitive and very cheap, but it is not at the top. One community verdict from launch day: certainly capable and very cost effective, but not better than the strongest models in raw capability.
Is it free anywhere?
The OpenRouter free launch period ended at 12:00 UTC on 22 October 2026 and the free variant was retired. Ollama showed the model as free until the same date. List price applies after that.
What else does MiniMax offer?
Video through the H3 family, speech through speech-2.8, images through image-01 — plus a M2.x family that is being retired. The model list has every identifier and every date.
Start with the
five-minute setup
Get a key, make one call, then wire it into the coding agent you already use.