MiniMax M3.1 Flash: API Price, Reasoning and M3 Migration
MiniMax M3.1 Flash is live with $0.15 input and $0.60 output per million tokens. See the exact API model ID, reasoning limits, cost math and M3 migration checks.
MiniMax released M3.1 Flash on October 19, 2026, pitching a faster, lower-cost model for coding agents, tool-using assistants and long-running business workflows. The practical question for developers is not whether the vendor's benchmark charts look impressive: it is how to call the model, what the published token prices imply for a real workload, and which limits still matter before switching an agent from MiniMax M3 or another provider.
What MiniMax actually released
MiniMax says M3.1 Flash is available through its API, M Plan and MiniMax Code. The official API documentation names the model MiniMax-M3.1-Flash and lists a 1,000,000-token context window. It supports text, image and video inputs, agent-oriented tool use and configurable reasoning depth. MiniMax M3 remains listed as a separate available model, so the new Flash release does not automatically mean every existing M3 integration has migrated.
Published M3.1 Flash API token rates
What the token price means in an agent workflow
Consider an illustrative batch with two million standard input tokens and half a million billed output tokens. At the published rates, that is $0.30 of input plus $0.30 of output, or $0.60 before any other charges or differences in actual metered usage. If one million of those input tokens instead qualifies for cached-input pricing, the arithmetic becomes $0.15 standard input + $0.015 cached input + $0.30 output = $0.465. These are billing examples, not measured task costs: an agent may make repeated calls, generate extra reasoning tokens or fail and retry.
The most useful comparison is cost per successful completed task, not dollars per million tokens alone. Log prompt, cached-input and output usage, tool calls, wall-clock time, retries and human correction for the same representative workflow on both models. A lower token price does not guarantee a cheaper result if an agent needs more steps or more supervision.
How M3.1 Flash differs from M3
Confirmed differences and what still needs testing
| MiniMax M3 | MiniMax M3.1 Flash |
|---|---|
| Existing model; still listed in official API documentation | New model launched October 19, 2026 |
| 1M context and multimodal inputs | 1M context and multimodal inputs |
| Existing integration may be stable for your workload | New configurable reasoning effort and published Flash rates |
| Do not assume a forced retirement | Benchmark, latency and task-cost improvements are vendor claims until tested on your work |
MiniMax says M3.1 Flash can run inference nearly three times faster and lower agent-task usage cost by 70% versus its predecessor. Those figures are the company's claims from its release materials, not independent guarantees for every API client, prompt or region. The official docs also specify that reasoning effort defaults to max; lower effort can reduce reasoning latency and output usage but may change results. Test accuracy and task completion before making low effort a production default.
API integration and a safe migration checklist
- Use the exact model identifier MiniMax-M3.1-Flash. The docs provide Anthropic-compatible and OpenAI-compatible API formats; match the correct base URL and effort parameter to your client.
- Run the same set of representative coding or business-agent tasks on M3 and M3.1 Flash. Measure completed tasks, token categories, total tool calls, retries and human review.
- Try low, medium and max reasoning effort on low-risk tasks. The documented choices also include high and xhigh; thinking cannot be switched off.
- Set a maximum cost and latency budget, preserve an M3 fallback, and review any outputs before granting an agent irreversible actions.
- Check the official model announcement and documentation again before planning self-hosting. The launch says model weights will arrive later, without a confirmed date.
The October 19 announcement says downloadable M3.1 Flash weights will be released later. Although a generic model-description line in the API documentation uses 'open-weight' wording, that is not evidence of a current public download, published license or release date. Treat self-hosting as unconfirmed until an actual weight release and license are available.
Questions people ask
Questions people ask
What model ID should I use for MiniMax M3.1 Flash in an API request?
Use MiniMax-M3.1-Flash, exactly as written in MiniMax's model invocation documentation. The model is shown for both Anthropic-compatible and OpenAI-compatible requests, but the endpoint and reasoning-effort field differ between protocols. Before replacing an existing M3 configuration, check your client library's endpoint settings and run a small test call rather than assuming that changing the model string is the only required change.
How much does MiniMax M3.1 Flash cost for a million tokens?
MiniMax's October 19 announcement lists $0.15 per million standard input tokens, $0.60 per million output tokens and $0.015 per million eligible cached-input tokens. These are separate billing categories, not a single flat cost for every million tokens. For an agent, calculate the actual mix of input, cache reads and output across all calls; retries and reasoning usage can change the cost per completed task.
Can I turn off reasoning in MiniMax M3.1 Flash to reduce latency?
No. MiniMax's official API documentation says thinking cannot be disabled for M3.1 Flash; requesting a disabled or none setting returns an error. You can instead reduce thinking depth using low, medium, high, xhigh or max effort. The default is max when effort is omitted. Lower settings may reduce latency and token use, but should be tested against task quality before being used in production.
Can I download and self-host MiniMax M3.1 Flash today?
A public weight download was not confirmed in MiniMax's October 19 launch announcement. The company said model weights would be released later but did not provide a release date or license there. One API documentation description calls the model open-weight, so the official pages are not fully aligned on availability. Do not plan a local deployment until MiniMax publishes an accessible weight package and applicable license.
M3.1 Flash is a credible new API option for cost-sensitive agents because MiniMax has published concrete pricing, a model identifier, a 1M context limit and reasoning controls. The migration decision should still be based on completed-task measurements rather than vendor benchmark or 70% savings claims. Keep M3 as a comparison baseline and treat local weights as pending until officially released.
Explore related Make Better resources
- What Should You Automate First?— Choosing workflows worth automating
- DeepSeek V4.1 Flash: API, Search Costs and Upgrade— Related agent model cost comparison; not the same product
Sources & useful resources
- MiniMax: M3.1 Flash launch, October 19, 2026— Official release, pricing and availability
- MiniMax API: Model Invocation— Official model identifier, context, effort and API integration
- MiniMax: M3 model launch— Official M3 context and product comparison