MiniMax M3.1 Flash Pricing and Cost Breakdown

Official token prices

MiniMax lists one flat rate for M3.1-Flash. There is no long-context tier to worry about, unlike the previous M3 generation, which doubles in price above 512k input tokens.

Two practical notes from the same pricing page:

• Token to word ratio. MiniMax estimates that roughly 750 English words consume 1,000 tokens. Use it as a rough budgeting rule, not a promise.
• Priority service tier. Passing service_tier: "priority" multiplies the standard rate by 1.5× for faster admission and better reliability. It is optional.

What the “70% cheaper” claim actually means

The launch post says M3.1 Flash delivers 70% lower overall usage cost for agentic tasks compared with the previous generation. That is a task-level cost claim, not a unit-price claim. It assumes a typical agentic workload — many turns, large prompts, tool calls — where reasoning output and repeated context dominate.

Two earlier, separate comparisons are worth keeping apart:

• Against M3. M3 costs $0.30/$1.20 per million input/output tokens up to 512k input, and $0.60/$2.40 above it. M3.1-Flash is cheaper at every size, and it is the cheaper option for long-context work because it does not step up.
• Against the Gemini and DeepSeek Flash tier. See M3.1 Flash vs DeepSeek V4.1 Flash for a same-run comparison, including the caveat that Artificial Analysis flags this model’s performance data as vendor-supplied.

Prompt caching is where the real savings are

Cached input costs $0.015 per million tokens — one tenth of the standard input rate, a 90% discount. Any agent loop that re-sends the same prefix (system prompt, tool definitions, a large file you are iterating on) should structure the request so that the stable part comes first and stays cacheable.

Two cost scenarios with the same 200k-token cached prefix:

That is roughly a 3× saving on a single call, and it repeats on every turn of an agent session.

What else costs money on the same platform

M3.1 Flash is the text model, but the same account bills the rest of the catalog. These are the published list rates.

Video — billed per second of output, not per token:

Audio input is free. The first 5 images per H3 task are free, then $0.04 per image; H3-Max includes the first 2 images and then charges $0.074 each. Input video is billed separately by duration and output resolution, and regenerating a video re-bills the original task’s input media.

Speech and image:

Voice cloning and voice design are charged when the voice is first used for synthesis, not when it is created. The full model list is on the model list page.

Cost per agentic task, measured and estimated

Artificial Analysis runs the same battery against many models, which makes its per-task numbers directly comparable across providers. For M3.1 Flash it reports a weighted cost per task of $0.22 and 615 seconds per task, built from about 78,600 output tokens per task (54,700 reasoning tokens plus 23,900 answer tokens).

That output volume is the thing to internalise: the reasoning budget dominates the bill. For a 100-task month at that average, roughly $22 of model spend, or $66 if your tasks are three times heavier than the benchmark mix.

Here is a rough estimate for three common shapes of work, so you can sanity-check your own invoice:

Every number in this table is our assumption about token counts, not an official figure. Replace the assumptions with your own usage before treating them as a budget.

The second cost lever: reasoning effort

The API accepts a thinking-effort setting with the values low, medium, high, xhigh and max. When you do not send the parameter, the default is max. If you are cost-mapping an existing workload, that single default can dominate everything else on this page — lower effort means fewer reasoning tokens for the same task.

Because thinking is always on for this model, you cannot simply switch reasoning off the way you could with M3. You can only change how deeply it thinks.

M Plan versus pay-as-you-go

MiniMax also sells subscriptions, which bundle the text model with image, audio and (on the higher tiers) video generation:

All modalities share one pool of usage, governed by a 5-hour window and a weekly window rather than a monthly bucket. Details are on the M Plan page.

Credit packs are the overflow mechanism: 1,000 Credits = $1, exactly the pay-as-you-go list price, valid for one year, spent only after your plan usage runs out and drawn from the pack that expires soonest. A credit pack is therefore not a discount — it is the same money in a wrapper that a Subscription Key is allowed to spend. (The mainland-China platform uses ¥7 for 1,000 Credits and lists Build at 12× instead of 7.5×.)

Rough break-even: if your monthly pay-as-you-go bill stays in the low tens of dollars, a flat subscription is likely to win on cost and remove the surprise-bill risk. If you run heavy batch jobs that only need the text model, pay-as-you-go at $0.15/$0.60 is hard to beat.

Free windows end quickly

Two free or discounted windows were live when this page was reviewed, and both are time-boxed:

• OpenRouter. The free launch period ended at 12:00 UTC on 22 October 2026, and the :free variant was retired. Requests now go to the base slug at MiniMax list price.
• Ollama. The library page still showed the model as free until 22 October 2026 for the cloud variant.

If you are reading this after those dates, treat the free tier as gone. The launch-date announcement and the current price page are the two sources that matter.

A price summary you can screenshot

• $0.15 input, $0.60 output, $0.015 cached input, per million tokens.
• No long-context step-up, unlike M3.
• Priority tier is 1.5×.
• Reasoning effort defaults to max — check it before you blame the token counter.
• Measured weighted cost per task: $0.22 on the Artificial Analysis battery.
• Subscription tiers run 1×, 3× and 7.5× relative usage.

Next: the API page shows how to set effort, caching and multimodal input, and the benchmarks page explains how much of these numbers to trust.