MiniMax M3.1 Flash pricing, plans and limits

Artificial Analysis gives M3.1 Flash a 43 on its Intelligence Index, and MiniMax's own table puts it ahead of Opus 5 on workflow automation while behind on terminal-style tasks. It suits teams running long tool-using agent loops on a small budget, and it is a poor pick where reasoning must be switched off or wrong guesses are costly.

MiniMax M3.1 Flash is a multimodal reasoning model with a 1,000,000-token context window that takes text, image and video input and returns text. MiniMax released it on 19 October 2026 as the agent-focused Flash tier of its M3 line, with always-on thinking, five effort settings and a Pro model promised later.

Provider: MiniMax · Family: MiniMax M3

More about MiniMax on HokAI

Context window: 1,000,000 tokens

Input modalities: text, image, video · Output: text

About MiniMax M3.1 Flash

MiniMax M3.1 Flash is a reasoning model from MiniMax, released on 19 October 2026 under the API name MiniMax-M3.1-Flash. It follows MiniMax M3, which shipped on 1 June 2026 after MiniMax M2.7, and MiniMax says an early build ran anonymously as the stealth model "Space Bunny" and processed more than 200 trillion tokens before the reveal. MiniMax's launch post and API docs give no parameter count, architecture note or technical report for the Flash variant, only a statement that "more aggressive sparsity" made inference faster. Press coverage that cites about 428 billion parameters is not backed by anything in those two documents.

On MiniMax's own comparison table, Flash scores 62.8 on Automation-Bench v1.0.6, a set of 600 public business-workflow tasks, against 50.3 for Opus 5 and 45.8 for GPT-5.6 Sol. It ties Claude Opus 5 at 76.2 on BankerToolBench, and scores 73.7 on DeepSWE v1.1 against 74.2 for DeepSeek-V4.1-Flash and 73.0 for GPT-5.6 Sol. The same table shows where it gives ground: 38.1 on Terminal Bench 4.0 against 51.8 for Opus 5, 61.7 on OSWorld 2.0 against 70.6, and 82.2 on CyberGym against 84.5 for GLM-5.3. These figures come from MiniMax's own table, not from an outside lab.

Independent numbers are thinner. Artificial Analysis scores Flash 43 on its Intelligence Index (version 4.3.2) and measured 122.8 output tokens per second with a 1.12 second time to first token on MiniMax's API. That speed is well under the 170+ tokens per second median MiniMax claims in its launch post, so plan capacity from the independent figure until more providers report.

Flash is the cheaper, always-reasoning sibling of M3. MiniMax reports agentic task cost 70% lower than its predecessor and inference nearly three times faster, both company figures, and says a Pro model will arrive later. Per token, Flash costs half of M3's standard rate, and unlike M3 it has a single price with no step-up for long prompts. Thinking is mandatory, so the saving shows up as cost per finished task, not as the cheapest possible single call.

Access today is through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Weights have been promised for later but no license or download exists yet, so the model sits in HokAI's MiniMax model directory as API-only until MiniMax publishes them. The directory files it under GA status, the 400K+ context bucket and proprietary openness, with image and video input and text output. The full model directory and the models hub list every comparable record. The documented rate limit is 200 requests per minute, the same as M3's.

Pricing

MiniMax lists one flat rate for MiniMax-M3.1-Flash: $0.15 per 1M input tokens, $0.60 per 1M output tokens and $0.015 per 1M cached input tokens, with no step-up for long prompts. Caching is automatic for requests of 512 or more input tokens, so agent loops that resend a long tool list and system prompt pay the cached rate on that prefix. The priority service tier that [MiniMax M3](/hub/models/minimax-m3) offers at a higher price is not supported on Flash. Because Flash always reasons, cost per task depends on the effort setting, which defaults to max when omitted.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• 1M-token context: Accepts a million tokens per request, enough for a mid-sized codebase or a long document set, at one flat rate with no long-prompt step-up.
• Tunable thinking depth: Five effort levels (low, medium, high, xhigh, max) control how much the model reasons before answering, and omitting the setting defaults to max.
• Text, image and video input: Chat requests can mix text, images and video, while output is text only.
• Automatic prompt caching: Repeated prefixes of 512 or more input tokens are cached without code changes and bill at a tenth of the input rate.
• Three API formats: Works through Anthropic-style messages, OpenAI Chat Completions and OpenAI Responses, so existing SDK code needs only a new base URL and model name.

Pros

• Cost per finished agent task is the selling point: Artificial Analysis recorded about $0.22 per Intelligence Index task.
• Ties Claude Opus 5 on BankerToolBench at 76.2 in MiniMax's own table, which suits document-heavy finance and operations workflows.
• A single price tier is easier to budget than M3, which steps up its price for long prompts.
• Leads every other model in MiniMax's comparison table on its Automation-Bench row, the closest match to SaaS workflow agents.

Cons

• Artificial Analysis recorded a 93% hallucination rate on AA-Omniscience, defined as wrong answers over all non-correct responses, so it tends to guess instead of declining.
• Very verbose: it produced 180M output tokens on the Intelligence Index run against a median of 99M, which lifts real cost per task above the headline rate.
• Reasoning cannot be turned off, so even trivial calls pay a latency and token floor.
• Trails Opus 5 and GPT-5.6 Sol on the terminal and desktop-control rows of MiniMax's own table.
• MiniMax's transparency page lists EU AI Act training-data summaries for M3 and H3 but not for Flash, and no system card is linked for it.

Benchmarks

• Humanity's Last Exam: 43.4% vendor-reported · 20 Oct 2026 — Expert-written questions across many fields, % correct.
• AA Intelligence Index: 43 cited: Artificial Analysis · 20 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 123 tok/s cited: Artificial Analysis · 20 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit MiniMax M3.1 Flash Official Page