MiniMax M3.1 Flash: Official Benchmarks, Cost, and Availability

One-sentence takeaway

MiniMax says M3.1 Flash is designed for Agent, coding, Computer Use, and cybersecurity tasks. The launch-page table lists DeepSWE v1.1 at 73.7%, OSWorld 2.0 at 61.7, and CyberGym at 82.2, with API pricing of $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached input tokens. These figures come from the vendor's published materials and should not be treated as independently remeasured results under a common harness.

Use cases

• Tasks this can help assess: Initial selection for tasks involving long tool-calling workflows, coding, enterprise workflows, Computer Use, or cybersecurity benchmarks; comparison of the relationship between publicly listed prices and officially reported performance/cost.

• Tasks this should not be extrapolated to: Treating official leaderboard scores as the accuracy, stability, or production success rate for every business; making strict cross-model rankings from results produced with different harnesses, model snapshots, or evaluation subsets.

• Applicable model version: MiniMax M3.1 Flash; the official API documentation lists the API ID as MiniMax-M3.1-Flash and describes support for a 1M context window and adjustable thinking depth.

• Test environment or client: The launch page uses different settings for different benchmarks; see “Evaluation method” and “Raw data.” This was not a single unified experiment. MiniMax says the model is available through the MiniMax API, M Plan, and MiniMax Code, and that the model weights will be released later.

• Reasoning tiers and parameters: The official charts show tiers including low, medium, high, xhigh, and max, but do not provide complete request parameters, temperature, sampling settings, repeat counts, or model snapshots. The Agents' Last Exam caption says each model uses its best harness and effort; M3.1 Flash is marked as Claude Code with max effort.

Evaluation method

This is an official launch material, with the body, tables, and charts all provided by MiniMax. The launch-page body summarizes vendor claims of “nearly three times the online inference speed” and “70% lower usage cost than the previous generation,” but it does not provide the previous-generation model, a complete speed-measurement protocol, or the cost calculation, so these statements can only be recorded as official claims.

The launch page's summary table places M3.1 Flash alongside other models, but its footnotes make clear that different benchmarks use different harnesses and evaluation rules. Some model columns are blank. MiniMax says a blank means there was no official evaluation or only a CLI evaluation was reported; a blank does not mean zero. The Agents' Last Exam caption also says that scores from different evaluation subsets or methods cannot be compared directly.

The chart publication dates also differ: the Automation-Bench and Agents' Last Exam captions mark the public data as current through 2026-10-18, while the DeepSWE caption marks it as current through 2026-09-22. The charts show M3.1 Flash's Automation-Bench scores by effort, but do not label the exact task cost for each point; billing figures cannot be estimated from the axes.

Key results

Results explicitly mentioned in the official body

• BankerToolBench: 76.2.

• DeepSWE v1.1: 73.7%.

• OSWorld V2: 61.7; Agents' Last Exam (ALE): 29.6. MiniMax says both are the highest among Chinese models, but the page does not provide the complete candidate set or unified remeasurement materials behind that claim.

• CyberGym: The body writes 82%; the benchmark table on the same page gives the precise value as 82.2. The table's precision should be used rather than turning the rounded body figure into a different result.

• Output speed: MiniMax says everyday use can reach 200+ TPS at peak, with P50 sustained at 170+ TPS; it does not publish input length, output length, concurrency, region, server version, or statistical-distribution details for the measurement.

M3.1 Flash column in the unified benchmark table

Pricing and availability

The launch page gives the following fixed prices:

MiniMax also says the overall usage cost for Agent tasks is 70% lower than the previous generation, but the page does not provide the previous-generation model, request composition, token ratio, or calculation sample. The claim therefore cannot be used to calculate actual savings for an arbitrary business.

Raw data

The complete M3.1 Flash column in the launch-page summary table is: AutomationBench v1.0.6 62.8, BankerToolBench 76.2, GDP.pdf AA 20.8, Terminal Bench 4.0 38.1, LOCA-Bench 256K 60.3, Agents' Last Exam 29.6, OSWorld 2.0 61.7, PostTrain Bench v1.1 39.3, DeepSWE v1.1 73.7, NL2Repo-Bench 61.2, SWE Refactor Bench 23.0, CyberGym 82.2, BabyVision (w/tools) 91.0, HLE (w/o tools) 43.4, and CharXiv RQ 82.5. Dashes in the table are missing values in other model columns, not results for M3.1 Flash.

The official launch page also shows three “score and cost” charts:

• The Automation-Bench chart lists M3.1 Flash's low/medium/high/xhigh/max scores as 52.8%/57.7%/59.2%/61.2%/62.8% and says task cost is plotted on a logarithmic scale; the exact cost for each M3.1 Flash tier is not labeled.

• The DeepSWE v1.1 chart shows M3.1 Flash at high 68.1%, xhigh 72.1%, and max 73.7%; the caption says the data is current through 2026-09-22, the evaluation harness is mini-swe-agent, and the chart shows average cost per task, but the M3.1 Flash points have no individual numeric labels.

• The Agents' Last Exam chart shows M3.1 Flash at max 29.6% with an estimated total cost of $111; the caption says ALE-V1 Overall has 152 tasks, each model uses its best harness and effort, and M3.1 Flash uses Claude Code with max effort. This cost is the estimated total cost under the chart's definition, not the charge for a single request in the API pricing table.

Conclusions and limitations

This launch page supports the following limited conclusions: M3.1 Flash is officially positioned as a low-cost Flash model for Agent tasks; MiniMax has published multiple benchmark scores, several harness configurations, and fixed token prices; and the official API documentation exposes the exact model ID MiniMax-M3.1-Flash. This is enough for initial selection and a reproduction plan.

The material cannot directly establish that:

• M3.1 Flash outperforms other models on any arbitrary business task. Different benchmarks use different harnesses, judges, context limits, and evaluation subsets.

• “Highest among Chinese models” or the cost/score relationships in the charts represent independently audited results. The material was published by the model provider, and the comparison conditions are not fully uniform.

• 200+ TPS, P50 170+ TPS, or 70% lower cost applies to every account, region, request length, and concurrency level. The page does not provide a complete measurement protocol or a previous-generation cost baseline.

• An API documentation description such as a 1M context window means that every deployment endpoint has the same effective context, latency, and tool limitations; these should be remeasured on the actual endpoint.

The launch page says the model is now available through the MiniMax API, M Plan, and MiniMax Code, and that the model weights will be released later. Availability is the official current statement; actual account permissions, region, plan, and client model list should be checked again at the time of use.

Discoverable independent evaluation entry points

The following links come from public benchmark entry points mentioned in the official charts or captions. They are leads for future independent verification and do not mean that these sites have already remeasured M3.1 Flash in this record:

• AutomationBench: The official caption says it uses 600 public tasks.

• Artificial Analysis: MiniMax says the GDP.pdf AA result is taken directly from this site.

• LOCA-Bench: The official caption points to its public implementation.

• Agents' Last Exam leaderboard: The official caption points to the ALE leaderboard.

• PostTrain Bench: The official caption points to this benchmark entry point.

• SWE Refactor Bench: The official caption uses this repository as the evaluation entry point.

Reproduction notes

For reproduction, fix and record: API ID MiniMax-M3.1-Flash, invocation entry point, account region and plan, model snapshot, thinking/effort tier, maximum context, tool and harness versions, request and output token counts, cache hits, concurrency, failed-request retries, cost, and evaluation date. For multiple benchmarks, use each benchmark's official implementation and scoring rules separately; do not combine scores from different harnesses into an overall accuracy rate.

To verify the launch page's cost claims, reproduce the API token prices, the per-task costs in the charts, and the estimated total cost for Agents' Last Exam separately, and record the input/output ratio for each. These three figures use different definitions and cannot substitute for one another.