MiniMax M3.1 Flash Review (2026): $0.15 Agents, No Weights

A model that spent three months as an anonymous stranger, processed 200 trillion tokens in public, and only then told everyone its name. The price is the pitch; the evidence is the problem.

By AI Tool Scout · Published October 20, 2026 · 17 min read

What Is MiniMax M3.1 Flash?

MiniMax M3.1 Flash is MiniMax's new low-cost agent model, released Monday, October 19, 2026, as a Flash-tier variant of MiniMax M3, which shipped in June 2026. It is the same family, aimed at a different buyer: M3 was the frontier-adjacent model, M3.1 Flash is the one you are supposed to run in a loop a thousand times a day. MiniMax's own framing is that it is redefining agent performance and cost, and the ordering of that sentence is honest — the cost claim is the sharp one, and the performance claim is the unverified one.

The specs are inherited rather than new, which matters for how you read the launch. 428 billion total parameters with roughly 23 billion active per token in a mixture-of-experts design, sparse attention to keep long sequences affordable, a native visual encoder rather than a bolted-on adapter, and a 1,000,000-token context. Input is text, image and video; output is text. MiniMax has not published a technical report describing what changed between M3 and the Flash variant, so the parameter figures are the family's, not a documented re-architecture, and anyone telling you otherwise is inferring.

What demonstrably changed is reasoning and price. Thinking is always on and cannot be disabled, with five effort levels — Low, Medium, High, Xhigh and Max — and MiniMax recommends Max as the default. Migration is not seamless: 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, instead of the inline <think> tags M3 emitted. On price, MiniMax halved M3's card to $0.15 input, $0.60 output and $0.015 cached input per million tokens, with no time-of-day windows, and claims that combination makes agentic tasks 70% cheaper and online inference nearly 3× faster than M3.

The company behind it is MiniMax, the Shanghai-based lab known outside China mainly for its video model Hailuo and its M-series text models, and it now ships on four surfaces at once: the MiniMax API, the M Plan subscription that replaced the old Token Plan, the MiniMax Code agent app (web, desktop, TUI, iOS and Android), and a promise that model weights will be released later alongside a stronger Pro model. That last line is the whole self-hosting story for now, and we will come back to it, because "later" is doing a lot of work in this launch.

The Space Bunny Story, and Why It Is the Best Thing in This Release

Before it had a name, this model had traffic. MiniMax's launch post says it plainly: Space Bunny, a model that operated anonymously and processed more than 200 trillion tokens, was an early version of MiniMax M3.1 Flash. Third-party coverage of the reveal adds the part MiniMax did not put in the headline — Space Bunny held the number one spot on OpenRouter for ten consecutive days and processed about 56.7 trillion tokens on that platform alone, which reporting describes as the highest volume ever recorded for a stealth model.

This is genuinely unusual, and it cuts both ways.

• The good reading: almost no model arrives with three months of real, adversarial, other-people's-workload traffic behind it. Anonymous models get tested by people who have no reason to be gentle — they run them on their own repos, their own support tickets, their own broken pipelines — and MiniMax says the tuning was driven by that. If you believe the tuning story, this is a model that has already survived a production burn-in that no benchmark replicates. It is a better signal than a benchmark table.
• The uncomfortable reading: 200 trillion tokens is not evidence of correctness. It is evidence of adoption — the model was fast and cheap enough that people kept coming back — and every one of those users was evaluating a model whose identity, terms, licence and data practices they could not see. That is a fact about the market, not a claim about MiniMax specifically; it is also the reason to discount "it was number one on OpenRouter" as a quality metric, because leaderboards on inference marketplaces rank volume, and volume follows price.
• The structural reading: the company had three months of feedback before it had a product page, and it still shipped no technical paper and no formal model card on day one. Whatever else the stealth period bought MiniMax, it did not buy documentation. Forkast's launch write-up makes the same point: the launch is opaque, and there has been no independent replication of the vendor's numbers.

So the honest summary of the Space Bunny affair is this: it is the most persuasive thing about M3.1 Flash, and it is not evidence you can audit. Keep it in the "why this is probably not a paper tiger" column and out of the "here is why it beats model X" column.

Key Features

• 428B / ~23B-active mixture of experts. Frontier-scale total capacity with a per-token compute bill closer to a 23B model — the whole economic basis of the release.
• 1,000,000-token context, flat-priced. No long-context surcharge anywhere on the card, so a 900K prompt costs the same per token as a 9K one.
• Text, image and video in; text out. The video input is the differentiated modality against most price-class rivals, and it comes with a native visual encoder rather than an adapter.
• Always-on reasoning with five effort levels. Low, Medium, High, Xhigh, Max, with Max recommended. You tune depth, not existence.
• Sparse attention. MiniMax's mechanism for keeping million-token sequences from becoming a quadratic bill.
• Speed: 200+ TPS at peak, P50 above 170. Enough that a coding agent's loop is limited by your tool calls, not by generation.
• Flat rate card. $0.15 / $0.60 / $0.015 per 1M input, output and cached input, with no peak or off-peak windows — the single most useful line in the pricing for anyone running 24/7.
• 90% cache discount. Cached input at a tenth of fresh input, which is the lever that decides whether a long agent loop is affordable.
• Three distribution routes plus a promise. MiniMax API, M Plan from about $22 a month, MiniMax Code on five platforms — and weights "later", with a stronger Pro model to follow.
• API-compatible with the tools you already run. M Plan issues a subscription key that plugs into Claude Code, Codex, Cursor, OpenCode and WorkBuddy, so the switching cost is a config change rather than an integration project.

Benchmarks: Five Numbers, One Vendor, Zero Replications

Here is MiniMax's table, reproduced as published. Read the last column twice.

All figures are vendor-reported as of October 19, 2026 on harnesses MiniMax selected. MiniMax says the model ranks first among Chinese models on both OSWorld V2 and Agents' Last Exam. It published no technical paper and no formal model card, and no independent reproduction existed at the time of writing.

Three things are worth saying about that table, and none of them are "these numbers are fake."

• Two of the six rows are not numbers. AutomationBench-AA is the benchmark most relevant to business buyers — real workflows through real APIs, which is what most "agent" projects actually are — and MiniMax says only that the model is "among the top performers." No score, no comparison table, no positioning. That is not a lie; it is an absence, and it happens to sit exactly where the purchase decision is.
• The one number with a rival to compare against puts it second. On DeepSWE v1.1 this model self-reports 73.7%; DeepSeek's V4.1-Flash-1017, released two days earlier at the same $0.15 / $0.60 headline, self-reports 77.0%. Both are vendor numbers on different harnesses, so neither debunks nor confirms the other, but if you were choosing on coding-agent claims alone, the cheaper comparison is not flattering to MiniMax. Our DeepSeek V4.1-Flash review has the other table in full.
• CyberGym at 82% is the most interesting number and the least checkable. Vulnerability discovery and patching is a workload where a confident wrong answer is expensive, and MiniMax's supporting evidence is that its internal security experts rated the model strong on code auditing, patching, penetration testing, malware analysis and reverse engineering. Internal expert review by the vendor is a real signal inside a company; it is not a benchmark, and it should not be quoted as one.

Pricing: the Flat Card Is the Product

Every figure below is approximate, in US dollars, checked on October 20, 2026. MiniMax's API rate card has no peak and off-peak split, which is the structural difference from DeepSeek and the reason this section can be short.

What a real agent task costs

Headline rates are useless without a token count, so here is one you can defend. Model a 30-step agent run — a coding or browser agent that calls a tool, reads the result, thinks and calls another — with a 6,000-token system prompt and tool schema re-sent every step, 1,200 tokens of tool output and 600 tokens of reasoning plus visible answer per step. Because each step re-sends the growing conversation, the run reads about 963,000 input tokens in total and writes about 18,000 output tokens.

Our arithmetic on the assumptions stated above, at published list prices. Competitor rates are approximate and checked October 20, 2026. Cache-hit share is an assumption; measure yours from your provider's dashboard before quoting any of this internally.

Two conclusions follow, and the first one is a correction to MiniMax's own marketing. At identical token counts the Flash card is 50% cheaper than M3, not 70% — the halved price accounts for all of it. The remaining gap to the claimed 70% has to come from the model using fewer tokens, which is what the "nearly 3× faster inference" and tuned reasoning-depth story is about. That is plausible and it is not something MiniMax published a token-per-task figure for, so treat 50% as the floor you can compute and 70% as the outcome you might observe. The second conclusion is that caching matters more than the headline price: the same run costs $0.16 uncached and $0.06 at 72% cache hits, which is a bigger swing than the difference between any two models on the table.

M Plan: the subscription route

MiniMax replaced its older Token Plan with M Plan at the same three price points, renamed and re-bundled. Monthly prices are confirmed by MiniMax's own platform pages; the platform's pricing table renders its numbers client-side, so confirm on the official page before you buy.

Three fine points that decide whether the subscription suits you. Yearly billing is ten months for twelve — the struck-through numbers are $264 / $648 / $1,560 — so the annual discount is about 17%, not a promotional gimmick. Usage is not published as a token count, only as a multiple of Go, so you cannot compute your own cost from it and should test your real workload inside a month before committing annually. And M Plan lists the text model as "M3.1 Flash Preview", while the API serves M3.1 Flash; if an exact model ID matters to your tooling, check what your account is actually routing to. The old Token Plan is closed to new buyers, with no M Plan equivalent for the per-seat Teams plan.

The first-month 50% discount that accompanied the M Plan launch applies to monthly billing only — $11, $27.50 or $66 for the first month, then full price. MiniMax's documentation listed the offer window as running through October 14, 2026, so as of today you should assume it has closed and verify on the checkout page rather than in a review written last month.

Is It Free?

No, and the distinction matters more here than usual because there are two different products under one name.

• The API has no free tier. No trial credit, no free request allowance, no zero-cost rate-limit band. Every token is billed from the first one at $0.15 / $0.60 / $0.015 per million. If your plan is to point Claude Code, Cursor or your own agent at MiniMax-M3.1-Flash, that plan has a price from call one.
• The consumer app does have a free column. MiniMax's own M Plan comparison table lists a Free tier: MiniMax Code on web, desktop, iOS and Android plus the CLI, a limited allowance of the M3.1 Flash Preview text model, limited image and speech models, and caps of 5 scheduled tasks, 20 website deployments and 20 projects. What it does not include is an API key — so it cannot be wired into your tooling — and, unsurprisingly, no video model. Sign in with a Google account and you can try the model today; the quota is a trial in practice, and MiniMax does not publish it as a number you can plan against.
• The 50% first month is a discount, not a free tier. It halves one month of monthly billing and then renews at full price. Annual billing is excluded from the offer.
• There is no open-weights escape hatch, yet. Every other "free" story in this price class runs through downloadable weights — DeepSeek's MIT checkpoint is free to run, forever, on your own hardware. MiniMax has not shipped weights, so the only zero-marginal-cost route is your own evaluation, not your own deployment.

Practical read: budget either ~$22/month for Go if one person is working in MiniMax Code, or pay-as-you-go if you are embedding the model in a pipeline. There is no third option where this model is both free and useful, and any listicle claiming otherwise is describing the consumer app's trial quota.

Pros & Cons

5 Best MiniMax M3.1 Flash Alternatives

Comparison Table

Prices approximate, in US dollars, checked October 20, 2026. A dash means the figure is not published on a comparable, current basis — not that it is zero. DeepSeek prices are off-peak first, peak second, and its peak windows are weekday-only in UTC. MiniMax's card has no peak split. Weight figures for MiniMax M3 are third-party reports, not a vendor-published checkpoint size, because no M3.1 Flash checkpoint exists to measure.

Which One Should You Choose?

• Running an agent pipeline at volume: shortlist it, pilot it, do not standardise on it yet. The flat card, the 90% cache discount, the 170+ TPS P50 and the 1M context are the right shape for loop-heavy work, and the price is competitive with DeepSeek's off-peak rate while being immune to its peak hours. But you are buying unverified benchmarks, so run your own eval on 50–100 real traces before you migrate a production path. If your traces show the model at or near DeepSeek's quality, the flat pricing wins the argument on a 24/7 workload; if they do not, nothing else in this review rescues it.
• Working a long codebase: it is a genuine option, but test recall at depth first. A 1M window with no long-context surcharge is rare, and video input means you can hand it a screen recording of a bug and nothing else. The catch is that MiniMax published no long-context evaluation, so the honest move is a retrieval test at the depth you actually need — 200K, 500K, 900K — rather than assuming the window is usable to the end. Compare against DeepSeek V4.1-Flash, which has a comparable window and an independent index score.
• Budget-sensitive: it is not the cheapest, and that is the point most reviews miss. At $0.15 / $0.60 it sits above GPT-6 Luna and Claude Haiku 5.5 at $0.10 / $0.50 and above GLM-5.3-Flash on output. What you are paying the premium for is depth — 428B parameters, always-on reasoning, 1M context, video in. If your workload does not need any of that, buy the $0.10 model and stop reading. If it does, $0.15 is still cheap, and the LLM API cost guide covers the levers — caching, output caps, effort levels — that matter more than the headline rate.
• Needing self-hosting: this is the wrong model today, and probably the wrong model after the weights land. No checkpoint, no licence, no size. When it arrives, the family's 428B / ~23B shape and roughly 440 GB of MXFP8 weights mean a multi-GPU node, so the realistic alternatives are DeepSeek's MIT checkpoint at 510 GB needing one 8× H200 node, an open family like Qwen that spans sizes you can actually run, or Kimi K3 if you have 1.56 TB and eight B300s. Choose MiniMax for the rate card, not for deployment control.
• Comparing before committing: use the free routes. MiniMax Code's free tier and DeepSeek's free chat both cost nothing, and both let you run the same twenty prompts through two models in an afternoon. That is worth more than any benchmark table in this article, including the one we reproduced.
• Buying the subscription rather than the API: do the arithmetic on the trial month first. Go at ~$22 is cheap enough to test, but M Plan publishes usage as a multiple of Go rather than a token count, so you cannot predict whether it covers your workload. Subscribe monthly, measure, then decide about the 17% annual discount.

Privacy & Data Note

MiniMax is a Shanghai-based company and the API is operated by it, which means the data-residency question is a jurisdiction question rather than a feature question. Three points belong in any risk review, and none of them are hypothetical.

• There is no self-host escape route today. With DeepSeek you can take the MIT weights and run them in your own VPC; here you cannot, because there are no weights. That removes the option most compliance teams reach for first, and it means "we will evaluate the model internally" is not a substitute for "we will not send client data to a Chinese-operated endpoint."
• Your prompts, tool outputs and files leave your perimeter. Agent workloads are the worst case for this: a coding agent reads source code, a finance agent reads ledger data, a browser agent reads whatever it can log into. Every one of those passes through the provider. If your data classification forbids that, the model is out regardless of price.
• Regional availability and terms change. Consumer surfaces and API access differ by country, and product availability, pricing and terms move quickly in this market. Confirm on MiniMax's official pages and with your own legal team what applies where you operate, and check whether a data-processing agreement is offered before you build a pipeline on top of it.

Two practical mitigations short of abandoning the model: keep sensitive material out of prompts and pass references rather than contents where your agent design allows it, and treat the consumer app and the API as separate decisions with separate risk profiles — the app's free tier is a great way to evaluate quality, and a bad place to paste anything you would not publish.

Frequently Asked Questions