MiniMax M3.1 Flash

Release Summary

MiniMax released M3.1 Flash on October 19, 2026, with hosted access through its API, M Plan and MiniMax Code. The main reasons to evaluate it against M3 are lower token rates and different reasoning controls. MiniMax also reports improvements in agent execution and inference speed; those are vendor results, not measurements from this update.

Weights are planned for October 23, 2026, according to MiniMax’s official announcement on X. That date is a release plan, not confirmation of downloadable weights, a particular license or working self-hosted deployment.

What Changed

Reduce token rates and remove M3’s long-context price step

The official API price table lists M3.1 Flash at $0.15 input, $0.60 output and $0.015 cached input per million tokens. Its flat rate contrasts with M3’s separate standard rates for inputs up to 512K tokens and above 512K tokens.

Against M3’s standard rate at up to 512K input, Flash halves the input and output token prices. Cache reads are one quarter of M3’s rate in that band. These rate comparisons do not establish a fixed saving per completed task: reasoning length, retries and cache hits also affect cost.

Retain the Preview's adjustable, always-on thinking

The earlier Preview already documented this behavior. The released M3.1 Flash retains five effort levels: low, medium, high, xhigh and max. Omitted effort defaults to max. M3 can answer with thinking disabled; Flash cannot, and an explicit disable request returns HTTP 400.

This changes the latency decision. Lower effort can reduce reasoning work, but it does not recreate M3’s non-thinking mode. An application that depends on that mode needs a separate evaluation before switching.

Improve agent execution and output throughput, according to MiniMax

MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam. It also reports output rates reaching 200+ tokens per second, with a 170+ TPS median. Those figures support evaluating Flash for coding agents and computer-use workflows; they do not guarantee the same results in another harness or at every effort level.

The company’s claim of 70% lower overall agent-task cost is distinct from the published per-token reductions. No same-harness M3-versus-Flash test was run for this update.

Use the released model ID in new integrations

The current SDK examples use MiniMax-M3.1-Flash, while the earlier Preview guide uses MiniMax-M3.1-Flash-Preview. Select the released ID explicitly when evaluating the final model; this update does not assume the Preview name is a permanent alias.

Compared With M3 (Previous Released Model)

This comparison uses MiniMax-M3 as the baseline for published API pricing and migration behavior. It is not a claim that the immediately preceding M3.1 Flash Preview lacked these capabilities. The Preview already offered 1M context, multimodal input and adjustable, always-on reasoning; those are retained features of the final release. The current SDK documentation lists both released models.

The 1M context and multimodal inputs were already available in M3. They are retained capabilities, rather than reasons on their own to upgrade. The price rows use M3’s standard tier, not its higher priority rates.

Compatibility & Migration

The released model ID is MiniMax-M3.1-Flash. The official documentation provides both Anthropic-compatible and OpenAI-compatible access, but changing only the model name can leave incompatible request settings.

• Remove requests that disable thinking, including thinking: {type: "disabled"} and effort none.
• On the Anthropic-compatible path, set effort through output_config.effort; on OpenAI-compatible Chat Completions, use reasoning_effort.
• Remove service_tier from Flash requests. Its absence of tier selection also matters if the M3 integration relied on priority admission.
• Preserve the full thinking and tool-use response history. OpenAI-compatible responses provide reasoning_content; Flash does not support forcing reasoning_split=false to embed thinking inside ordinary content.
• Budget for reasoning tokens. The Anthropic-compatible documentation states that thinking counts toward max_tokens; too small a budget can end the response before answer text appears.

M Plan subscription access and pay-as-you-go API billing are separate access paths. The token prices above describe the latter; they do not specify a subscription allowance or prove that an older Preview entitlement has changed on every account.

Availability & Planned Weights

The launch announcement confirms hosted availability through the MiniMax API, M Plan and MiniMax Code. It says weights will follow; the official X post supplies the more specific October 23 plan.

As checked on October 20, 2026, this page treats the weights as planned. A self-hosting decision still needs an actual official checkpoint, its license and verified serving instructions. It does not assume parameter counts, GPU memory requirements or deployment support from M3 or the Preview.

The MiniMax tool overview explains the product’s broader access options. This release page focuses on what changes for someone evaluating M3.1 Flash.

Who Should Upgrade / Who Should Wait

Evaluate an upgrade if API token spend is a material part of an existing M3 coding or tool-use workflow, especially when requests cross M3’s 512K pricing boundary. Flash also merits evaluation when the application can use adjustable thinking and its parser already preserves reasoning and tool-call history.

Wait or keep the existing M3 route if the application needs a non-thinking response mode, depends on M3 priority admission, or requires locally available weights and a confirmed deployment license today. A lower token rate alone does not resolve those constraints.

Sources

Last verified: October 20, 2026. Source-based research from the official product materials below. No hands-on testing was performed for this update.

• M3.1 Flash release announcement — October 19
• Official X announcement — planned October 23 weights
• Current API pricing — Flash and M3 bands
• Anthropic-compatible API — models, thinking and compatibility
• OpenAI-compatible API — effort and reasoning response handling
• Model overview — context and multimodality
• M3 announcement — previous-version baseline
• Earlier M3.1 Flash Preview — retained thinking controls