MiniMax M3.1 Flash and the New Price of an AI Task

What MiniMax Released

MiniMax made M3.1 Flash generally available on October 19, 2026, through the MiniMax API, its M Plan subscription and its MiniMax Code tool. Developers had already met it. Per MarkTechPost and Unite.ai, it first appeared on September 27 as an anonymous OpenRouter model called “Space Bunny” and as M3.1-Flash-Preview inside MiniMax Code.

MarkTechPost reports a mixture-of-experts design with 428 billion total parameters and about 23 billion active per token, a 1-million-token context window and multimodal input. MiniMax hasn’t published a technical report on Flash, so treat the parameter count as reported, not confirmed.

The price is the headline. MiniMax lists $0.15 per million input tokens and $0.60 per million output tokens, and says agentic tasks cost about 70% less than on its predecessor, MiniMax M3. Artificial Analysis’s figures point the same way, if less steeply: M3 scores 29 on the v4.3 index at $0.51 per task, while M3.1 Flash’s first listing is 42 at about $0.22. That is 13 more points for roughly 57% less per task.

The Cost-per-Task Ladder

Artificial Analysis runs each model through the same set of evaluations, records the tokens it uses, and multiplies them by the model’s list price. The result is a dollar figure per task. The table below puts M3.1 Flash’s first listing next to the v4.3 figures for the models a buyer is most likely to weigh it against, sorted by score.

Source: Artificial Analysis Intelligence Index v4.3.2, except MiniMax M3.1 Flash, which is a single first listing that we haven’t been able to cross-check. Gemini 4 Argon’s figure uses a 50%-off launch price.

Start with the four models at 42. M3.1 Flash is the cheapest of them at about $0.22 per task, a few cents under the MIT-licensed GLM-5.3-Flash at $0.25. Claude Opus 5.5 at low effort costs 2.5 times as much for the same score, and GPT-5.6 Terra about six times as much.

Then look one tier up. MiMo-V2.6-Pro, open-weight under the MIT license, scores four points higher at about $0.13 per task. So the accurate claim is a narrow one. M3.1 Flash is the cheapest listed model per task at its capability level. It is not the cheapest model per task, and it isn’t the best value on the board. If the score-42 tier is good enough for your work, there is a new cheapest option. If you want the most capability per dollar, MiMo-V2.6-Pro still comes first.

Treat $0.22 as a launch-day reading. First listings can shift as providers adjust and as Artificial Analysis re-runs its evaluations, and a few cents is all that separates M3.1 Flash from GLM-5.3-Flash.

Cheap Tokens, Long Answers

Per token, M3.1 Flash is very cheap. Its input price is about one twenty-seventh of Claude Opus 5.5’s $4 per million, and its output price about one thirty-third of Opus 5.5’s $20. Yet on Artificial Analysis’s numbers it is only 2.5 times cheaper per task than Opus 5.5 at low effort. A token discount of roughly 30 times shrinks to a task discount of 2.5 times. The difference is how many tokens each model spends to finish.

Artificial Analysis describes M3.1 Flash as very verbose. Across the whole index it produced 180 million output tokens, against a median of 99 million, or about 1.8 times as much text as the typical model. For Opus 5.5 at low effort to land at $0.55 per task with prices that much higher, it has to finish the same work with far fewer tokens.

The practical lesson is simple. The per-token price on a rate card is where the calculation starts, not the answer. A model that writes twice as much at half the token price costs the same per task. The cheap rate is real, but some of it goes to paying for extra words.

MiniMax’s own chart gives lower figures: under $0.05 per task on AutomationBench at every effort level, and about $0.28 per task on DeepSWE. Those are MiniMax’s measurements on its own task sets, so they can’t be compared with Artificial Analysis’s $0.22. The spread between them shows how much cost per task depends on the kind of task.

How to Read Cost per Task

Cost per task is the most useful price figure in AI right now, and the easiest to misread. Six rules cover most of it.

1. It is price times tokens — The figure multiplies the list price by the input and output tokens a model actually used on the benchmark, so it captures verbosity that a rate card hides.
2. It belongs to one benchmark — Artificial Analysis’s number is the cost of its own index tasks. Your prompts, document lengths and retries will differ, so use it to rank models, not to forecast an invoice.
3. Check the index version — Artificial Analysis rescaled its index to v4.3 on September 7, 2026. Scores fell and costs rose, so a v4.1 figure can’t sit next to a v4.3 one.
4. Effort settings are separate entries — Claude Opus 5.5 is listed at $0.55 per task on low effort (42), $3.46 on xhigh (56) and $5.98 on max (58). Same model, more than a tenfold spread.
5. Promotions and first listings move — Gemini 4 Argon’s $1.99 depends on a 50%-off launch price, and M3.1 Flash’s $0.22 is a first listing.
6. It assumes a per-token meter — On hardware you own there is no list price to multiply. The cost becomes the hardware, the power and the time each task takes.

For a longer treatment of the metric and how the v4.3 rescale changed it, see Dollars per Task.

MiniMax’s Benchmarks Against DeepSeek and GLM

MiniMax compared M3.1 Flash with two open-weight rivals: DeepSeek’s October 17 update, V4.1 Flash 1017, and Z.ai’s GLM-5.3. Every score in this table comes from MiniMax. None has been reproduced independently.

By MiniMax’s own scoring, M3.1 Flash trails DeepSeek V4.1 Flash on both coding-agent tests and leads it on AutomationBench at max effort. At low effort its AutomationBench score falls below DeepSeek’s. It beats GLM-5.3 on all three, though only by 0.2 points on Terminal-Bench 4.0. The rival figures match what DeepSeek and Z.ai report for their own models (77.0 and 42.3 for DeepSeek, 66.9 on DeepSWE for GLM-5.3), so the comparison is at least consistent.

The row that matters most for buyers is the price row. DeepSeek’s off-peak rate is the same $0.15 and $0.60 that MiniMax charges, and DeepSeek V4.1 Flash already ships MIT-licensed weights. That is the comparison M3.1 Flash has to win. Our sources don’t include an Artificial Analysis cost per task for DeepSeek V4.1 Flash, so we can’t place it on the ladder above.

Open Weights on October 23, License Not Yet Announced

MiniMax says the M3.1 Flash weights will be published on October 23, 2026. It hasn’t said under what license. That gap matters. “Open weights” means the files can be downloaded. The license decides whether you can use them commercially, what you have to display, and whether you need anyone’s permission first.

The predecessor is the best guide to what MiniMax tends to ask for, but it isn’t a promise. MiniMax M3’s weights ship under a custom “minimax-community” license.

If M3.1 Flash follows M3, a company with under $20 million in yearly revenue would send MiniMax a one-time notice and show a “Built with” line. A company above that line would need written authorization before using it. Most custom open-weight licenses aim their conditions at very large model-as-a-service providers and leave internal use by ordinary businesses alone. Z.ai’s GLM-5.3 license, for example, adds a security review only for model-as-a-service businesses above $10 billion in annual revenue. M3’s $20 million line applies to organizations in general, so it can catch an ordinary mid-sized firm.

Licenses are moving in both directions. Z.ai went from MIT on GLM-5.2 to a custom license on GLM-5.3, and Moonshot moved from Kimi K2’s Modified MIT to a custom Kimi K3 license. Google went the other way, releasing Gemma 4 under Apache 2.0. Nobody outside MiniMax knows which way M3.1 Flash will go.

The practical advice is to wait four days. Don’t plan a self-hosted deployment of M3.1 Flash until the license text is out and someone has read it. If you need certainty today at the same score, GLM-5.3-Flash also scores 42 and is MIT-licensed. This is not legal advice; check the license text. We track the terms model by model in the Open-Weight AI License Tracker.

What It Would Take to Self-Host

If MarkTechPost’s 428-billion-parameter figure holds, M3.1 Flash needs roughly 240 GB for its weights at 4-bit quantization. That is the same footprint as MiniMax M3, which has the same reported size. It fits a three-unit Mac Studio M5 Max configuration with 384 GB of unified memory.

Faraday Machines installs open-weight models on Mac Studio hardware in Toronto. Our Scale tier is that three-unit, 384 GB configuration at $29,999 US, and MiniMax M3 is already among the models we install on it. Whether M3.1 Flash joins it depends on the weights actually shipping on October 23, the license allowing your use, and testing on real hardware. We won’t make claims about its speed until it has been measured.

Running it yourself changes what verbosity costs. With no per-token meter, M3.1 Flash’s extra output no longer shows up on an invoice. It shows up as time instead, because a model that writes 1.8 times as much text takes longer to finish each answer on any hardware. The bill is fixed, and the prompts and documents stay in your building. For the wider picture of which open models fit which hardware, see Open Source AI Models Compared.

References