MiniMax's fast, low-cost agentic model — output above 200 tokens per second, flat $0.15/$0.60 per-million-token pricing, tool use, coding, and image input, with a context window reported around one million tokens.
Last verified · 2026-10-19 · by Moe Ameen
MiniMax M3.1 Flash is the fast, low-cost tier of MiniMax's M-series, rolled out on October 19, 2026. You reach it through the MiniMax API, the M Plan, and MiniMax Code, where it had been in preview since late September; MiniMax has confirmed that the model benchmarked anonymously as "Space Bunny," which processed more than 200 trillion tokens, was an early version. It is a reasoning model built for agentic work — tool use, coding, computer use, and multi-step automation — rather than casual chat.
Two numbers define it: speed and price. MiniMax cites everyday output above 200 tokens per second (a median near 170) and roughly three times faster inference than before, and prices it flat at $0.15 per million input tokens and $0.60 per million output, with cached input at $0.015 and — per third-party reporting — no long-context surcharge. On agentic tasks it claims the cost lands about 70% below its predecessor. It reads text and images and returns text, and trackers list a context window around one million tokens, which is what makes it practical to drop a long source in and get structure back.
MiniMax's own numbers lead on agents and coding — 76.2 on BankerToolBench, 73.7% on DeepSWE, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, 82% on CyberGym — and it claims to top Chinese models on OSWorld V2 and ALE. Independent trackers are more measured: Artificial Analysis scores it 42 on its intelligence index (a jump over MiniMax M3, level with GLM-5.3-Flash, behind DeepSeek V4.1 Flash 1017) and reports a high hallucination rate, so lean on it for drafting and structure and verify any facts.
The boundary that matters for content work is simple: M3.1 Flash reads text and images and writes text. It generates no image, video, caption, carousel, or post, keeps no persona's face or voice consistent, and publishes nowhere. MiniMax has not confirmed a parameter count and says open weights and a stronger "Pro" model will come later — so treat specifics as provisional and confirm them against MiniMax's own page.
The thing M3.1 Flash is unusually good at is reading something long, fast. A million-token context and 200-tokens-a-second output mean you can hand it a full webinar transcript, a long report, or a month of notes and get back a clean set of angles, outlines, and extracted quotes in seconds, for cents. What it can't do is turn any of that into something your audience actually sees. That next move — structure becomes finished, published media — is exactly what [Kompozy](/) does, and it is built to take that kind of distilled brief as its input.
Here is the loop. Feed your long source to M3.1 Flash and ask for the angles worth posting — the ten strongest points, a hook per point, a short script for each. Bring that shortlist into Kompozy and it fans each angle into a format the model can't render: a captioned [Persona Shorts](/glossary/persona-shorts) avatar video narrating the script, a brand-exact [Carousel Post](/glossary/hyperframes) of the key points, a Quote Graphic of the sharpest line, native Text Posts written per platform, and a [Blog Article](/ai-tools/koala-ai) and Email Newsletter from the long version. A single [Persona Brief](/glossary/persona-brief) with banned-word filters holds one voice across the whole set — and because the model's weak spot is accuracy, Kompozy's per-post review gate is where you catch anything it got wrong before it ships. Then [Autopilot](/glossary/autopilot) schedules and publishes across the eight social platforms plus blog and email. (Kompozy's own copy runs on Claude and OpenAI, so M3.1 Flash is an upstream distillation-and-drafting tool, not a plug-in.) The model turns a long source into a brief; Kompozy turns the brief into a published week.
It is MiniMax's fast, low-cost agentic model, rolled out on October 19, 2026 via the MiniMax API, M Plan, and MiniMax Code. It is built for tool use, coding, and multi-step automation, reads text and images, returns text, and lists a context window reported around one million tokens. MiniMax says it runs about three times faster than its previous model.
MiniMax lists $0.15 per million input tokens and $0.60 per million output, with cached input at $0.015 per million, charged flat with — per third-party reporting — no long-context surcharge. MiniMax also claims roughly 70% lower cost on agentic tasks than its predecessor. Confirm current rates on MiniMax's site.
It is fast and cheap for drafting and for distilling long sources, but independent testing flags a high hallucination rate, so treat its output as a draft to fact-check rather than finished copy. Use it for structure and angles, and keep a human review step on anything factual.
No. It generates text only — no images, video, captions, carousels, or audio — and it publishes nothing. To turn its drafts into finished, multi-platform content, pair it with a generation-and-publishing engine like Kompozy.
Distill a long source or draft a batch on the model, then bring the shortlist into Kompozy. Kompozy generates captioned video, carousels, quote graphics, photo posts, blogs, and newsletters in your brand voice from those seeds and publishes them across the eight social platforms plus blog and email.