MiniMax reveals M3.1 Flash, with low API prices and uneven agent results

The API is here; the weights have a promised date

MiniMax lists $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input. Tokens are the chunks of information processed by a model; cached input is previously processed material reused at a discounted rate. These are API unit prices, rather than a subscription price or a fixed bill for completing a job.

The current pricing table makes one comparison straightforward: Flash’s input and output rates are half those of MiniMax-M3 for requests up to 512,000 input tokens. MiniMax’s broader claim of a 70% reduction in agent usage cost also depends on how much work and how many tokens a run requires.

The original announcement below promises model weights on 23 October. The developer guide already describes an open-weight model, but that wording does not establish that the files and their licence are available today. Space Bunny being an earlier version also means old preview experiences should not automatically be treated as tests of this release.

Strong workflow results do not cover every kind of work

Artificial Analysis says it tested M3.1 Flash at maximum reasoning effort. Its 19 October report puts the weighted average cost at about $0.22 per Intelligence Index task, compared with about $0.51 for MiniMax-M3. That is an independent result for that evaluation mix, not a quote for building a website or running a campaign.

The task breakdown is more useful than a single winner label. Artificial Analysis reports a rounded 73% on AutomationBench-AA, but 26% on Terminal-Bench v4.0. The first measures the share of workflow objectives achieved without guardrail violations in simulated business applications; it is not a 73% rate of finishing whole customer jobs. Its live methodology page shows the unrounded 72.5% result.

The tools around a model matter too. Our earlier DeepSeek coding-agent report shows how changing the agent software and reasoning setting changes cost and time for the same checkpoint. MiniMax’s guide says Flash’s thinking cannot be switched off, and that omitting the effort setting selects its maximum level. Lower unit prices leave that configuration decision very much alive.

A landing page is a useful example, with a limited verdict

Hassan (@nutlope), whose profile identifies him as head of developer experience at Together AI, posted a side-by-side landing-page demonstration for a hardware launch. He reports a $0.014 MiniMax run against $1.22 for Claude Opus 5.5, alongside a speed advantage. His video below shows two visibly different page designs.

This is a concrete demonstration from someone working at an inference provider. The inspected post does not disclose the full prompt, agent software, reasoning settings, repeated runs or an independent quality assessment. Its bill comparison therefore cannot establish an equally good result at an 88-fold saving across design work. A cheap first draft and a finished page are different deliverables.

The knowledge test exposes a separate trade-off

Artificial Analysis’s knowledge test supplies the strongest reason to keep those distinctions. On AA-Omniscience, its 19 October report gives Flash 56% correct answers while attempting 98% of questions. It reports a 93% hallucination rate among non-correct outcomes. That last figure is conditional, not a claim that 93% of all answers are wrong.

The benchmark’s method counts incorrect answers against the pool of wrong, partial and unanswered outcomes. A model can therefore know more answers while becoming much less willing to abstain when it does not know. That matters when an agent fills a page with product facts: visual polish does not verify the copy.

The release makes M3.1 Flash worth evaluating for cost-sensitive agent work, with the evidence strongest for the tasks actually tested. Compare completed, checked deliverables under matched settings. Treat 23 October as a promised weight-release date, and assess the files, licence and hardware requirements when they arrive.