MiniMax M3.1 Flash: 70% Cheaper AI API With 1M Context

MiniMax M3.1 Flash: 70% Cheaper AI API With 1M Context

Two days ago, a Chinese AI startup did something the big labs usually avoid: it launched a genuinely capable coding and agent model, gave it a one-million-token memory, and priced it so low it undercuts almost everything on the market. MiniMax M3.1 Flash went live on October 19, 2026 — and at $0.15 per million input tokens, it just fired the next shot in the AI price war.

If you build with AI APIs — coding agents, document pipelines, anything that burns tokens at scale — this one is worth your attention. Here is what M3.1 Flash actually is, what it costs, how it stacks up against DeepSeek and the rest, and how you can try it today.

⚡ Quick answer: MiniMax M3.1 Flash is a multimodal AI model for coding and agent tasks with a 1M-token context window, launched October 19, 2026. It costs $0.15 per million input tokens and $0.60 per million output tokens — flat, with no long-context surcharge — and MiniMax says it runs agentic workloads at about 70% lower cost than its predecessor M3. Open model weights are promised for October 23, 2026.

Table of contents

What is MiniMax M3.1 Flash?

MiniMax is a Shanghai-based AI company behind the M-series language models. M3.1 Flash is the fast, lightweight variant of that family — built specifically for coding, computer use, and multi-step agent work rather than as a general-purpose chatbot flagship.

The headline specs, as reported by the company:

  • Architecture: a mixture-of-experts model MiniMax describes as roughly 430 billion total parameters, with only about 23 billion active per token — which is why it can run fast and cheap.
  • Context window: 1,000,000 tokens — enough to hold an entire codebase, a stack of long documents, or a very long agent conversation in one session.
  • Multimodal input: it takes text, images, and video as input. It can read a screenshot, a screen recording of an error, or a photo of a whiteboard, and it answers in text. For comparison, Google's Nano Banana Pro showed how far image-understanding models have come — but M3.1 Flash is aimed at builders, not image generation.
  • Adjustable reasoning: an effort parameter with five levels (low, medium, high, xhigh, max). Max is the default and thinking can't be switched off entirely — but if you want faster, cheaper answers, you just lower the level.
  • Speed: MiniMax claims online inference is about 3x faster than M3, with 200+ output tokens per second in everyday use.

Fun detail: the model first surfaced as a stealth model codenamed "Space Bunny" before MiniMax revealed it as M3.1 Flash — the kind of quiet test-run that has become standard practice before a public launch.

The price is the real story

Here's the part that matters most. MiniMax M3.1 Flash is priced at:

  • $0.15 per million input tokens
  • $0.60 per million output tokens
  • $0.015 per million cached input tokens — a 90% discount when the model re-reads context it has already seen

There is no long-context surcharge — you pay the same rate whether you use 10,000 tokens or the full million. That detail is a bigger deal than it sounds: several providers quietly double their rates past certain context thresholds. MiniMax doesn't.

How does that compare? Here's the current "cheap model" leaderboard:

Model Input / 1M tokens Output / 1M tokens Context Weights
MiniMax M3.1 Flash$0.15$0.601MDue Oct 23 (license TBA)
DeepSeek V4.1 Flash$0.30$1.201MOpen (MIT)
GLM-5.3$1.40$4.40~1MOpen weights
Kimi K3$3.00$15.00~1MOpen (Kimi K3 License)

Read that pricing column again. MiniMax is half the price of DeepSeek's flash model and one-twentieth the price of Kimi K3 per input token. MiniMax itself frames the comparison differently — it says agentic tasks cost about 70% less than on its own predecessor, M3, which it describes as one-third the cost and three times the speed.

This is part of a wider trend we've been tracking in our AI statistics roundup: frontier-capable intelligence keeps getting cheaper, and the "Flash" tier — fast, cheap, good-enough models — is where the real competition now lives.

How good is it, actually?

Cheap means nothing if the model can't do the job. Here is what the benchmarks say — with one important caveat: these are MiniMax-reported numbers, and independent third-party evaluations haven't landed yet. Treat vendor benchmarks as a signal, not gospel.

  • Automation-Bench v1.0.6: 62.2 — ahead of DeepSeek V4.1 Flash (54.8) in MiniMax's own comparison table
  • DeepSWE v1.1: 73.7 — essentially tied with DeepSeek V4.1 Flash (74.2)
  • Terminal Bench 4.0: 38.1 — ahead of DeepSeek V4.1 Flash (31.2)
  • BabyVision (with tools): 91.0 — ahead of DeepSeek (89.6)
  • CyberGym (security): 82% — which MiniMax says shows leading ability in vulnerability discovery, code auditing, and patching

The honest read: on agentic and automation benchmarks, M3.1 Flash looks genuinely competitive with — and in places ahead of — its closest rival, DeepSeek V4.1 Flash. But the frontier flagships still sit above it: in MiniMax's own table, Claude Opus 5 scores higher on Terminal Bench, OSWorld, and several coding tests, and GPT-5.6 Sol leads on document-heavy benchmarks. M3.1 Flash isn't trying to be the smartest model alive. It's trying to be the cheapest model that's smart enough — and on that metric, the early numbers look strong.

MiniMax M3.1 Flash vs DeepSeek V4.1 Flash

This is the comparison everyone will make, because DeepSeek V4.1 Flash has been the budget-king pick since its launch. Here's how they line up:

  • Price: MiniMax wins clearly — $0.15/$0.60 vs DeepSeek's $0.30/$1.20 (peak rates; DeepSeek offers half off-peak).
  • Context: Tie — both offer 1M tokens.
  • Openness: DeepSeek wins today — V4.1 Flash weights are already out under the MIT license. MiniMax's weights are promised for October 23, with the license still undisclosed.
  • Benchmarks: Near-tie with a slight MiniMax edge on automation tasks (per MiniMax's own numbers).
  • Ecosystem: DeepSeek has the bigger community and tooling footprint right now; MiniMax counters with OpenAI-compatible and Anthropic-compatible API endpoints, so switching costs are low either way.

Bottom line: if you want the cheapest capable API right now and don't need to self-host, MiniMax M3.1 Flash is the new price leader. If you need open weights under a permissive license today, DeepSeek V4.1 Flash keeps its crown — at least until October 23.

How to try MiniMax M3.1 Flash (step by step)

Good news: you don't have to wait. The model is already live. Here's how to get hands-on:

  1. Via the MiniMax API (cheapest): sign up at platform.minimax.io, grab an API key, and call the model as MiniMax-M3.1-Flash. Endpoints are OpenAI-compatible and Anthropic-compatible, so your existing code needs minimal changes.
  2. Via OpenRouter (easiest test drive): OpenRouter ran a 72-hour free window at launch, and now lists the model at the same $0.15/$0.60 rates — handy if you already route through OpenRouter.
  3. Via M Plan: MiniMax's subscription tier gives a fixed monthly usage allowance at a flat price — worth it if your usage is steady rather than spiky.
  4. Via MiniMax Code: the company's desktop coding assistant lets you pick M3.1 Flash for bug-fixing and function-writing directly in the app.
  5. Self-hosting (from Oct 23): wait for the weight release, check the license terms carefully, then deploy with your usual inference stack.

One practical tip: because cached input costs just $0.015 per million tokens, design your agents to reuse context — keep the same system prompt and document blocks across calls instead of rebuilding them. That's where the real savings compound.

And if you end up recording a demo or tutorial video of your M3.1 Flash build for YouTube, you can grab a clean, high-res thumbnail from any video in one click with our free YouTube Thumbnail Downloader — handy for the video's own thumbnail too.

Who should use it — and who shouldn't

M3.1 Flash is for you if:

  • You run coding agents or computer-use agents and your API bill keeps you up at night
  • You process long documents or codebases and need a real 1M-token context without a surcharge
  • You want a drop-in cheap model via OpenAI-compatible endpoints
  • You're prototyping agent workflows and want maximum experimentation per dollar

Look elsewhere if:

  • You need fully open weights with a known permissive license right now — DeepSeek V4.1 Flash (MIT) is the safer pick until October 23
  • You're doing frontier-level reasoning, complex math, or novel research — the flagship tier (Claude Opus 5, GPT-5.6-class) still leads there
  • Your compliance setup requires a disclosed license and audited model card before touching a vendor model

The catch: what we still don't know

Let's be honest about the gaps, because the launch coverage mostly isn't:

  • The license is undisclosed. "Weights on October 23" sounds great until you learn whether it's MIT, Apache, a custom research-only license, or something restrictive. Don't plan your self-hosting roadmap until you see the terms.
  • The benchmarks are vendor-reported. MiniMax's table is flattering to MiniMax — shocking, we know. Independent evals from the community will be the real test, and they take a few weeks.
  • The "Space Bunny" preview was limited. The preview build was restricted to M Plan and MiniMax Code users, so public track records are thin.
  • Chinese-model considerations. Like other China-built models, expect the usual questions about data handling and content policies — check MiniMax's terms against your own requirements, especially for enterprise use.

Pros and cons

✅ Pros

  • Cheapest capable API in its class — $0.15/$0.60 per million tokens
  • Full 1M-token context with no long-context surcharge
  • Strong agent/coding benchmarks (vendor-reported) vs DeepSeek V4.1 Flash
  • Multimodal input: text, images, and video
  • OpenAI + Anthropic-compatible endpoints make switching easy
  • Adjustable reasoning effort to trade smarts for speed

❌ Cons

  • Weights not yet released; license unknown until October 23
  • All headline benchmarks are company-reported, not independently verified
  • Not a frontier reasoning model — flagships still win on hard tasks
  • Smaller ecosystem and community than DeepSeek or the US labs

FAQs

How much does MiniMax M3.1 Flash cost?

MiniMax lists M3.1 Flash at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. There is no extra charge for using the full 1M-token context.

When will MiniMax M3.1 Flash weights be released?

MiniMax says the model weights will arrive on October 23, 2026. The license type has not been disclosed yet, so check the terms before planning a self-hosted deployment.

Is MiniMax M3.1 Flash better than DeepSeek V4.1 Flash?

On price, yes — it costs about half as much per token. On MiniMax's own benchmarks it edges DeepSeek on automation tasks and ties on coding. But DeepSeek's weights are already open under MIT, while MiniMax's are still pending.

How do I access MiniMax M3.1 Flash?

Through the MiniMax API at platform.minimax.io (model ID MiniMax-M3.1-Flash), via OpenRouter, through the M Plan subscription, or inside the MiniMax Code desktop app. The API supports OpenAI-compatible and Anthropic-compatible endpoints.

What is the context window of MiniMax M3.1 Flash?

Up to 1,000,000 tokens, with no long-context price surcharge. It accepts text, image, and video inputs and returns text.

What was "Space Bunny"?

"Space Bunny" was the stealth codename MiniMax used while quietly testing M3.1 Flash before the public launch on October 19, 2026. The preview was limited to M Plan subscribers and MiniMax Code users.

Verdict

MiniMax M3.1 Flash is the most aggressive pricing move in AI this month — a genuinely capable 1M-context agent model at half of DeepSeek's flash pricing, with no long-context tax and OpenAI-compatible access. If your workload is token-hungry agent loops, long-document processing, or coding assistance, it's the new default budget pick.

The one thing to watch is October 23: if the weights drop under a permissive license, M3.1 Flash goes from "cheapest good API" to "cheapest good model you can run yourself" — and the price war gets even more interesting. We'll update this post when the license lands.

Last updated: October 21, 2026 · Written by The AI Tools team (about us).

How we researched this: we cross-checked the launch facts across multiple independent outlets — MarkTechPost's spec breakdown, Unite.AI's launch coverage, and Tech Insider's pricing table — against MiniMax's own API documentation figures, and compared pricing claims with the public rate cards of DeepSeek, Moonshot AI, and Zhipu. Benchmark figures quoted are MiniMax-reported and flagged as such; independent evaluations were not yet available at publication time. No sponsorships, no affiliate links — if we got something wrong, tell us and we'll fix it. Sources: MarkTechPost, Unite.AI, Tech Insider, Medium / Data Science in Your Pocket.


Comments

Post a Comment