MiniMax M3.1 Flash explained: agent benchmarks, pricing, and whether to upgrade

The takeaway

MiniMax released M3.1 Flash on October 19, 2026. The central question is whether it can complete longer agent workflows at a lower cost: understand the goal, call tools, operate software, check results, and continue.

For teams already using MiniMax, a controlled trial on real work is the sensible next step. Official scores are useful evidence of potential; they do not establish that every existing workflow will become faster or cheaper.

This is an analysis of official release material. TboxIO has not independently rerun these benchmarks or measured API latency. Information checked on October 20, 2026.

What changed?

The announcement emphasizes complex automation, software engineering, computer use, and security tasks. Agent quality depends on an entire chain of actions. A good answer is only the start: consistency across files, reliable tool calls, and recovery from mistakes determine whether the task gets delivered.

For workflows involving business applications, spreadsheets, files, or repositories, evaluate the completed outcome rather than the fluency of the explanation.

Official benchmarks: what each score measures

These are vendor-reported results. The benchmarks use different tasks and scales and should not be averaged into one score.

Benchmark	Official M3.1 Flash result	Main task area
BankerToolBench	76.2	Multi-step planning, file consistency and tool use
DeepSWE v1.1	73.7%	Understanding code, implementing changes and validating them
OSWorld V2	61.7	Computer operation and cross-application work
Agents’ Last Exam (ALE)	29.6	Complex professional tasks
CyberGym	82%	Vulnerability discovery and validation

DeepSWE is a more relevant starting point for repository maintenance; OSWorld and ALE are more relevant to desktop and professional workflows. None of these scores is a forecast of the success rate on your own tasks. Tools, step limits, retries, prompts, and grading can change the outcome. Security benchmarks also do not replace authorized testing and human review of a particular system.

What can we compare with M3?

The M3 announcement, dated June 1, 2026, emphasized long context, multimodality, coding, and agents. M3.1 Flash puts more emphasis on execution efficiency and usage cost.

Two comparisons need care:

• M3 reported benchmarks including SWE-Bench Pro and Terminal-Bench 2.1, while the new announcement highlights DeepSWE v1.1. These are different tests; subtracting their scores does not measure a coding improvement.
• M3’s methodology describes OSWorld-Verified and the effect of changing the maximum step count. The new release reports OSWorld V2. A higher or lower number alone does not establish progress or regression.

Quantifying an improvement over M3 requires the same dataset version, agent framework, and budget. We do not invent missing historical scores or convert incompatible results into a percentage gain.

Pricing and speed in practical terms

The release announcement lists the following API prices in USD per million tokens:

Item	Announced price
Input	$0.15
Output	$0.60
Cached input	$0.015

For example, a batch with two million uncached input tokens and 500,000 output tokens would cost approximately $0.60 in model token fees: $0.30 for input plus $0.30 for output. This illustration excludes other service fees; actual charges depend on billed usage.

MiniMax claims a 70% reduction in overall agent-task usage cost versus the predecessor and nearly threefold faster online inference. It also reports output rates reaching 200+ tokens per second, with a P50 of 170+ tokens per second. These are vendor claims, not guarantees for each request.

For a real project, cost per accepted completed task matters more than token price alone. Retries, large tool responses, and manual rework can erase savings. Output throughput is also different from end-to-end task latency, which includes queuing, tool execution, and external services.

Who should try it first?

Teams with repeatable workflows: start with reversible tasks whose outputs are easy to check, such as file organization, report drafts, or proposed code changes.

Developers constrained by cost or throughput: retain the current model as a baseline and compare identical inputs, permissions, and acceptance criteria.

Computer-use users: check software versions, interface languages, and permissions. A benchmark desktop does not represent every business application. Keep explicit confirmation for consequential submissions, payments, and deletion.

A useful pre-release comparison

1. Select historical tasks, including ordinary cases, long workflows, and known failures.
2. Fix the environment, prompts, step limits, and retry budget for both models.
3. Record completion rate, manual edits, end-to-end time, and input, output, and cached-token usage.
4. Check whether failures stop safely, especially when files or multiple systems are modified.
5. Increase traffic only when outputs meet your acceptance criteria.

This tests the efficiency claims while exposing business-specific problems that public leaderboards may miss.

Availability and model weights

The announcement lists MiniMax API, M Plan, and MiniMax Code as access routes. It says weights will be released later. API availability should therefore not be described as weights already being downloadable; verify the official repository, license, and runtime requirements before local deployment.

Our judgment: the most useful test of M3.1 Flash is whether completion quality and cost improve together on long workflows. Validate that combination on your own work before making a broad migration.

Sources and comparison scope

• Official M3.1 Flash release, October 19, 2026: reported scores, speed, release-time pricing and access routes.
• Official M3 release, June 1, 2026: predecessor positioning and evaluation methods, not a basis for subtracting incompatible scores.

Prices, availability, and weight-release status can change. Check current official information when integrating or deploying.