MiniMax M3.1 Flash landed on 19 October 2026 with a blunt pitch: near-frontier agentic scores at $0.15 per million input tokens — and open weights arriving on 23 October. This guide shows you how to try it today through the API, M Plan, MiniMax Code or OpenRouter, what it actually costs per task, and where MiniMax’s marketing claims diverge from what independent testers measured.
What MiniMax M3.1 Flash actually is
M3.1 Flash is MiniMax’s efficiency-first variant of its M3 line, built for agentic work, coding and visual reasoning. The numbers that matter, all from MiniMax’s own product page and launch coverage: 428 billion total parameters with about 23 billion active (mixture-of-experts), a 1,000,000-token context window, and multimodal input — text, image and video — with text output. It uses sparse attention and a native visual encoder, and reasoning is always on: you pick one of five effort levels (Low, Medium, High, Xhigh, Max), with Max as the recommended default.
One migration gotcha is worth knowing up front. Requests that send thinking.type=disabled or effort=none now return an error, and reasoning arrives in a separate reasoning_content field that may be absent — so integrations that parsed inline think tags from M3 need updating before they work with Flash.
There is also a fun provenance detail: “Space Bunny”, the anonymous model that quietly processed more than 200 trillion tokens on public leaderboards, was confirmed by MiniMax as an earlier M3.1 Flash build. If you liked Space Bunny’s leaderboard run, you were already benchmarking this model.
How to try it today: four access paths
You don’t need to wait for the weights — M3.1 Flash is live right now on four surfaces. Here is each path and who it suits:
| Path | What it is | Best for |
|---|---|---|
| MiniMax API | Direct API at platform.minimax.io, flat per-token pricing | Developers wiring agents into existing pipelines |
| M Plan | MiniMax’s subscription route to the model | Steady, non-API users who want predictable billing |
| MiniMax Code | Desktop coding app with the model built in | Trying the model on real coding tasks fast |
| OpenRouter | Listed after a 72-hour free period, then flat $0.15/$0.60/$0.015 pricing | Zero-setup trials and side-by-side model comparisons |
The OpenRouter route is the cheapest way to kick the tyres: the 72-hour free window lets you run real agentic tasks before spending anything, and OpenRouter’s unified pricing afterwards matches MiniMax’s own.
What it actually costs: the per-task arithmetic
The headline price is $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015 — but the number that matters for agent workloads is cost per task. Artificial Analysis’s per-task measurement puts M3.1 Flash at about $0.22 per task, roughly three times faster than the older M3. That per-task figure is the honest yardstick, because agent runs burn wildly different token mixes.
How does that sit against the other cheap-agent contenders we have covered?
| Model | Input / output per 1M | AA Intelligence Index | Catch |
|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 / $0.60 | 42 (MiniMax-reported; 13 pts above M3) | All capability benchmarks vendor-reported so far |
| DeepSeek V4.1-Flash-1017 | $0.15 / $0.60 | — | Identical list price; MIT weights already out — see our DeepSeek V4.1-Flash run guide |
| Claude Haiku 5.5 | $0.10 / $0.50 (under 100K prompts) | 43 | 5x price cliff past 100K-token prompts — see our Haiku 5.5 review |
The interesting bit: M3.1 Flash and DeepSeek V4.1-Flash share identical list prices, so the choice between them comes down to measured per-task cost and which vendor’s benchmark claims you trust — which brings us to the honesty split.
Vendor claims vs independent measurements
MiniMax’s launch claims are aggressive; the independent measurements that exist so far are good but smaller. Quote the measured numbers, not the marketing ones:
| Claim | MiniMax says | Independent says | Status |
|---|---|---|---|
| Cheaper than M3 | 70% lower cost on agentic tasks | Artificial Analysis measured ~57% per-task reduction | Real saving, overstated |
| Speed | Over 200 tokens/sec everyday use, P50 above 170 | Artificial Analysis measured 120 tokens/sec | Fast, not 200 |
| Intelligence | AA Index 42, first among Chinese models on Automation-Bench | No independent index score yet | Unverified |
| Coding | DeepSWE 73.7%, Terminal Bench 4.0 38.1 | No independent runs yet | Vendor-only; mid-pack vs Opus 5’s 51.8 on Terminal Bench |
This is the same vendor-vs-independent pattern we apply in our testing methodology: vendor benchmarks are directional, not verdicts. M3.1 Flash still lands at the top of the cheap-agent bracket on measured cost — a ~57% per-task cut and $0.22 per task is genuinely strong — but “70% cheaper” and “200 tokens a second” are the numbers to retire from your notes.
Getting ready for the 23 October weights
MiniMax said on X that the open weights arrive on 23 October 2026 — though its own blog hedged with “later”, so treat the date as stated-not-guaranteed. If you plan to self-host, run this five-point readiness check this week:
- Watch for the licence first. No licence terms have been disclosed yet. Don’t commit hardware or compliance sign-off until the licence is public — this is the single biggest unknown.
- Don’t size hardware yet. MiniMax hasn’t disclosed self-hosting requirements. At 428B total / ~23B active, this will not be a single-GPU model, but anything more specific is guesswork until release day.
- Use the official repo only. When the weights land, download from MiniMax’s official release channel — not mirrors or re-uploads, which are where tampered weights circulate.
- Budget the reasoning migration. The always-on reasoning change (no disabled mode, separate reasoning_content field) affects self-hosted inference stacks too — plan the integration update now.
- Trial on the API first. Use the OpenRouter free window or the MiniMax API this week to learn whether the model’s agentic behaviour fits your workload before you spend money on hardware.
Our take: who should use it now, who should wait
Try M3.1 Flash today if you run budget-sensitive agent workloads; wait if you need self-hosting or verified hard-coding performance. The desk-researched verdict, with the honest caveat that we have not hands-on tested this model:
- Agent builders on a budget: try it this week. ~$0.22 per measured task at 1M context is the strongest cheap-agent value on the board right now, and the OpenRouter free window makes the trial free.
- Teams already on DeepSeek V4.1-Flash: don’t switch on price alone — list prices are identical. Run your own per-task comparison first; the difference will be output-token frugality on your specific workload.
- Self-hosters: wait for 23 October and the licence. No licence, no hardware specs, no commitment.
- Teams buying for hard coding: wait for independent benchmarks. MiniMax’s own table puts it mid-pack on Terminal Bench 4.0 (38.1 vs Opus 5’s 51.8), and every coding score is vendor-reported.
MiniMax also says a larger Pro model is coming later with no date — Flash is explicitly the efficiency play, not the flagship. Judge it on cost per completed task, and it currently wins its bracket. Judge it on verified frontier coding, and the jury is still out.
Frequently asked questions
Is MiniMax M3.1 Flash free to try?
Partially. OpenRouter listed the model after a 72-hour free period, so there was a free trial window at launch; ongoing use is pay-per-token at $0.15/$0.60 per million input/output tokens. The MiniMax API, M Plan and MiniMax Code are paid surfaces.
When do the MiniMax M3.1 Flash open weights release?
MiniMax said on X that weights arrive on 23 October 2026, though the company blog only committed to “later”. Treat 23 October as the stated date, and wait for the licence terms — none have been disclosed yet — before planning a self-hosted deployment.
Is M3.1 Flash really 70% cheaper than M3?
That is MiniMax’s claim for agentic-task cost. Artificial Analysis independently measured roughly a 57% per-task reduction — still a large saving, but meaningfully smaller than the marketed figure. We recommend quoting the measured number.
How is M3.1 Flash different from the regular M3?
Same 428B-total / ~23B-active architecture, but tuned as the efficiency variant: MiniMax claims about 3x the inference speed of M3, always-on reasoning with five effort levels, and the lower $0.15/$0.60 pricing. A larger Pro model is promised later.
Can M3.1 Flash replace Claude or GPT models for coding?
Not on current evidence. MiniMax’s own benchmarks show mid-pack hard-coding scores (Terminal Bench 4.0: 38.1 vs Opus 5’s 51.8), and no independent coding benchmarks exist yet. It is a cost play for agentic work, not a verified frontier coder.
Sources and methodology
Claims rest on launch-week reporting: MarkTechPost’s 19 October release breakdown (specs, pricing, reasoning behaviour), Unite.AI’s launch coverage (benchmarks, availability, weights date), Oossa’s per-task cost analysis (~$0.22/task, AA Index 42), AI News Daily’s X-announcement recap (70% claim, Space Bunny confirmation), and a 20 October independent review noting Artificial Analysis’s 120 tokens/sec measurement. All capability benchmarks are MiniMax-reported unless labelled independent; vendor cost/speed claims are shown beside independent measurements. Verification date: 20 October 2026. Pricing and availability describe the 19 October launch and may change. This guide is desk-researched, not hands-on tested.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.