MiniMax M3.1 Flash: The $0.15 Agent-Task Model

Updated: October 2026 · 6 min read · By the CreatorSprout team

MiniMax M3.1 Flash abstract illustration — Abstract illustration created for this article

Launch coverage based on MiniMax's release materials and press reporting — we have not tested this model hands-on. All benchmark scores below are MiniMax's own figures unless marked as independent; treat manufacturer claims as unverified. Pricing verified October 2026; confirm on official pages before you commit. No affiliate links in this article.

The headline for builders: flat pricing with no peak-hour surcharge, a claimed 70% drop in per-task agent costs, and open weights scheduled for October 23, 2026. Here's what the launch actually contains — and where the marketing claims part company with independent data.

What M3.1 Flash is

M3.1 Flash is MiniMax's lightweight counterpart to its flagship models — built for coding, terminal work, and agentic tasks where speed and cost per task matter more than maximum reasoning depth. Size barely changed from M3: MiniMax's developer materials list 428 billion total parameters with roughly 23 billion active per token (a sparse mixture-of-experts design), a 1-million-token context window, and text, image, and video input.

The extra speed comes, per MiniMax, from "more aggressive sparsity" at inference time — the company claims nearly three times the inference speed of M3 and output above 200 tokens per second in everyday use. One architectural quirk to know: reasoning is always on and cannot be disabled. Developers choose one of five effort levels — Low, Medium, High, Xhigh, Max — with Max recommended as the default. Requests that try to switch thinking off return an error, so teams migrating from M3 will need to update any integrations that parsed inline thinking tags.

The benchmarks (mostly MiniMax's own numbers)

Every score below is MiniMax's self-reported figure unless labeled independent. None have been independently re-checked, and MiniMax hasn't published a technical report on what changed in the Flash variant:

DeepSWE v1.1 (real-world software engineering): 73.7% — just below DeepSeek V4.1-Flash-1017's 77%
OSWorld V2 (operating a computer): 61.7 — MiniMax says no Chinese model has scored higher
BankerToolBench (banking workflows): 76.2
CyberGym (security work): 82% — code auditing, pentesting, malware analysis
AutomationBench v1.0.6: 62.2 — the top score in MiniMax's own comparison table
Humanity's Last Exam: 48% — ahead of DeepSeek V4.1 Flash's 43% and GLM-5.3-Flash
AA-Omniscience (factual accuracy): 56% — versus 43% for DeepSeek V4.1 Flash
Agents' Last Exam: 29.6%
Terminal Bench 4.0: 38.1 — a weak spot; one review notes Anthropic's Opus 5 scores 51.8 here

The independent picture is thinner but telling. Artificial Analysis placed M3.1 Flash at 42–43 on its Intelligence Index (reports cite both 42 and 43 from the same index), behind DeepSeek V4.1 Flash 1017's 45, and measured 120 tokens per second — well below MiniMax's 200+ claim. Artificial Analysis also measured the average task cost at $0.22 versus $0.51 for M3 — a 57% reduction, real but short of the advertised 70%.

Pricing: flat $0.15/$0.60 — and the DeepSeek comparison

M3.1 Flash API pricing is $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. The flat structure is the differentiator:

No peak-hour pricing. MiniMax charges the same rates around the clock — unlike DeepSeek V4.1 Flash, which matches $0.15/$0.60 off-peak but doubles to $0.30/$1.20 during weekday peak hours.
No long-context surcharge. M3 doubled its rate above 512,000 input tokens; M3.1 Flash runs the full 1M-token window at the listed price.
Cached input is pricier than DeepSeek's. $0.015 per million versus DeepSeek's $0.003 — a 5x gap that matters for long-running agents that lean on cache.

So: the sticker prices match DeepSeek exactly, MiniMax wins on predictable flat rates, DeepSeek wins on cached-token cost and (slightly) on independent benchmark scores. Pick based on which workload you actually run.

Open weights coming October 23

MiniMax says the model weights will be released on October 23, 2026. As of this article, the license terms and self-hosting hardware requirements haven't been announced — so don't assume MIT or Apache-style terms until MiniMax actually publishes them. MiniMax also teased a "Pro" model that "pushes the boundaries of intelligence," with no date attached.

The "Space Bunny" backstory

Here's the fun part of the launch story: MiniMax confirmed that "Space Bunny" — a mystery model that had been running anonymously on public benchmark leaderboards and processed over 200 trillion tokens — was an early M3.1 Flash build. It worked: Space Bunny built a reputation before anyone knew who it was, and the reveal turned a spec sheet into a story. Take the 200-trillion-tokens figure as MiniMax's claim, but the strategy is confirmed by the company itself.

How to try it

MiniMax API: live now at the company's developer platform; MiniMax's M Plan subscription is the other API route.
MiniMax Code: MiniMax's desktop app, aimed at developers.
OpenRouter: listed at matching prices; one review noted it was offered free for its first 72 hours — check whether that's still live.
Self-host: not yet — weights land October 23, license still unannounced.

One thing to check if you already use M3: prompts and integrations tuned around the old model's thinking format may break, since reasoning is now always-on with the effort-level system. Test before assuming parity.

The honest caveats

The 70% claim is MiniMax's own math. The company hasn't disclosed how it measured it; independent measurement on a standard index landed at 57%. Different workloads, different numbers — but the headline figure is unverified.
Independent scores trail DeepSeek. 42–43 vs 45 on the Intelligence Index, and the speed gap between the 200+ tokens/sec claim and the 120 measured. The "generational leap" framing is MiniMax's, not the data's.
Reasoning can't be switched off. Five effort levels is the control you get. Integrations parsing inline thinking output from M3 will need rework.
Weights aren't out yet. October 23 is the promise; license terms and hardware requirements are still unknown as of publication.
Terminal coding is a weak spot. Terminal Bench 4.0 at 38.1 lags the field — relevant if your agents live in the CLI.

Verdict

MiniMax M3.1 Flash is the honest value play of October's model-launch window — not the most capable new model, but arguably the most practical one for high-volume agent work. Flat $0.15/$0.60 pricing with no peak surcharge is genuinely simpler to budget than DeepSeek's two-tier schedule, and even the independently measured 57% task-cost reduction is a real number. The trade: slightly weaker independent scores than DeepSeek's V4.1 Flash, always-on reasoning you'll have to design around, and weights you still can't download yet. If your agents run around the clock, the flat pricing alone may make M3.1 Flash your default router choice — just run your own evals before you commit production traffic.

FAQ

What is MiniMax M3.1 Flash?

A model update from Chinese AI lab MiniMax, released October 19, 2026. It's a 428-billion-parameter mixture-of-experts model (about 23B active per token) with a 1-million-token context window, built for coding, computer use, and agentic tasks. MiniMax calls it a faster, cheaper successor to the M3 model it launched in June.

How much does MiniMax M3.1 Flash cost?

$0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens — flat rates with no peak-hour pricing and no long-context surcharge. MiniMax claims agentic tasks cost 70% less than on M3; independent measurement put the reduction at 57%.

How does it compare to DeepSeek V4.1 Flash?

The two launched three days apart at the same $0.15/$0.60 sticker price. M3.1 Flash has flat rates; DeepSeek doubles to $0.30/$1.20 in peak hours but charges far less for cached input ($0.003 vs $0.015). On independent scores, DeepSeek's build edges it — 45 vs 42–43 on the Artificial Analysis Intelligence Index, and 77% vs 73.7% on DeepSWE v1.1.

Is M3.1 Flash open source? Can I self-host?

Not yet. MiniMax says the weights will be released on October 23, 2026, but hasn't announced the license terms or hardware requirements. Don't assume permissive terms until the release actually lands.

Was "Space Bunny" really M3.1 Flash?

Yes — MiniMax confirmed that Space Bunny, a model that ran anonymously on benchmark leaderboards before the announcement, was an early M3.1 Flash build. MiniMax claims it processed over 200 trillion tokens in that period.