MiniMax M3.1 Flash Explained: 2026 Price & Speed Guide

On October 19, 2026, Chinese AI startup MiniMax launched M3.1 Flash, a refreshed version of its M3 model family built for one specific job: running AI agents fast and cheap. The company promises a "generational" jump in agentic tasks, programming, Computer Use and cybersecurity — with inference nearly three times faster and usage costs down 70 percent.

For developers and teams building agent products, that combination is worth paying attention to. But it is also worth separating what MiniMax actually announced from what remains unverified. This guide covers both.

What MiniMax actually launched

M3.1 Flash is a multimodal mixture-of-experts (MoE) model with 428 billion total parameters and about 23 billion active parameters, served through the MiniMax API with a 1,000,000-token context window. It accepts text, image and video input and returns text, and it targets coding, agentic office work and visual reasoning.

The model is live now on three surfaces: the MiniMax API, the company's M Plan subscription tier, and MiniMax Code, its coding-focused product line. Notably, MiniMax confirmed that the mysterious "Space Bunny" model spotted in the wild was actually an earlier M3.1 Flash build — a stealth test run that the company says processed over 200 trillion tokens before the official announcement.

Model weights are scheduled for public release on October 23, 2026, which would let researchers and developers download and run the model outside MiniMax's hosted API.

The pricing: $0.15 in, $0.60 out, $0.015 cached

The most concrete numbers in the announcement are the API prices:

| Model | Input (per 1M tokens) | Output (per 1M tokens) | Cached input (per 1M) | Context | |---|---|---|---|---| | MiniMax M3.1 Flash | $0.15 | $0.60 | $0.015 | 1M tokens |

Because the model uses fewer output tokens per task, MiniMax estimates the effective cost at roughly $0.22 per agentic task.

The cached-input price deserves special attention: at $0.015 per million tokens, it is a 90 percent discount versus fresh input. That pricing is explicitly designed for agentic workflows where a model repeatedly reasons over the same long context — a codebase, a document set, or a running conversation history. It is the same playbook DeepSeek V4.1 Flash and Mistral Large 4 have used to compete for developer mindshare: win on price-per-task, not on flagship benchmark crowns.

Benchmarks — all company-reported

MiniMax published a set of benchmark scores alongside the launch. Every figure below comes from the company itself or from outlets relaying MiniMax's numbers; independent verification is not yet available:

• Humanity's Last Exam: 48% for M3.1 Flash, versus 43% for DeepSeek V4.1 Flash and below GLM-5.3-Flash; MiniMax says it ranks first among Chinese models.
• Artificial Analysis Intelligence Index: 42 points, 13 points higher than the previous M3.
• CyberGym (cybersecurity): 82%, which MiniMax presents as leading capability in vulnerability discovery and validation.
• Automation-Bench v1.0.6: 62.2, the top score in MiniMax's own comparison table.
• Terminal Bench 4.0: 38.1, notably behind Claude Opus 5's 51.8 — an honest-looking weak spot in MiniMax's own table.
• Output speed: median (P50) above 170 tokens per second in everyday use, per MiniMax's measurements.

That mix of strength (factual accuracy, agentic office tasks) and weakness (hard terminal coding) is actually useful information. It suggests M3.1 Flash is a strong candidate for document-heavy, multi-step agent loops and a weaker pick for autonomous terminal work — at least until third parties can confirm.

Five things developers should know

1. Reasoning cannot be turned off. M3.1 Flash always reasons; requests that set thinking to disabled return an error. You choose from five effort levels — Low, Medium, High, Xhigh and Max, with Max as the default. If you integrated the earlier M3 model and parse inline thinking tags, plan to update your code, since reasoning now arrives in a separate field that may be absent.

2. Thinking tokens count toward the output limit. Keep this in mind for jobs with tight response caps: a long reasoning chain eats into your max-token budget.

3. Multi-turn tool calls need full history. For tool-use loops, you must retain the complete assistant response in the conversation history to preserve reasoning continuity across turns.

4. The 1M context is input and output combined. The million-token ceiling is shared, not one million each way — the same limit the previous M3 had, so this launch does not expand context capacity.

5. Weights come on Oct 23. Until then, nobody outside MiniMax can run the model locally or audit its behavior independently. If you evaluate it now, you are evaluating the hosted API only.

Why this launch matters

The AI pricing war is now the industry's dominant dynamic. When a "flash-class" model delivers near-frontier agentic scores at $0.15 per million input tokens, the economic floor for running autonomous agents drops for everyone — including startups that could never afford frontier-tier pricing. That is good news for builders, and it keeps pressure on every other lab to match or beat these rates.

But there is a healthy way to read this announcement and a careless way. The careless way treats MiniMax's "3x faster" and "generational improvement" claims as verified facts. The healthy way treats them as a strong starting hypothesis and runs a pilot: put M3.1 Flash and your current model side by side on your own most common task, measure task-completion quality and actual cost per completed task, and then decide. For Computer Use deployments in particular — where a model can click through your desktop — that controlled pilot is a security practice, not just an economic one.

MiniMax also said a Pro model is coming later, described as pushing the boundaries of intelligence further. Flash is the efficiency play; Pro would be the capability play. Watching how the two are positioned will tell you a lot about where MiniMax thinks the market splits.

The honest bottom line

M3.1 Flash is a real, available model with aggressive pricing, a generous context window, and a clear target: high-volume agentic and coding workloads. The headline numbers are compelling, but they are company-reported, and the model's weakest published score (Terminal Bench 4.0) is a reminder that "flash" models make trade-offs. Try it cheaply, benchmark it yourself, and let your own workload — not the launch post — decide.

Sources: Reuters via TrendinTech's full breakdown · MarkTechPost · Unite.AI · Oossa · Tech Insider