MiniMax announced the general rollout of M3.1 Flash on October 19, 2026, positioning the model around agentic work, software development, computer use, and cybersecurity. The company says online inference is nearly three times faster than its predecessor and that the overall usage cost for agentic tasks is 70% lower. Those are company-reported claims, not independent guarantees: real speed and task cost depend on the workload, prompts, tools, output length, and evaluation method.
For developers, the practical news is that MiniMax says M3.1 Flash is available through the MiniMax API, M Plan, and MiniMax Code. Its published token rates are $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. The model’s weights are not yet publicly released; MiniMax says they will follow later.
This guide separates confirmed availability and published prices from performance claims that still need independent testing, and explains what developers should evaluate before choosing M3.1 Flash for a coding agent or automated workflow.
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is a model designed for practical, multi-step work rather than only single-turn chat. MiniMax highlights agentic tasks, coding, computer use, and cybersecurity as its main areas of improvement. In an agent workflow, a model may need to interpret a goal, plan steps, call tools, inspect results, recover from errors, and continue until a task is complete.
That distinction matters because a model that produces a quick answer is not automatically an effective agent. For software development, the useful outcome may be a working patch that passes tests. For computer use, it may be a sequence of actions that correctly completes a task across applications. For business automation, it may be a completed workflow with the right records updated and the result verified.
MiniMax says M3.1 Flash is a new iteration of its Flash line and that an earlier version operated anonymously as “Space Bunny,” processing more than 200 trillion tokens. That background is part of the company’s launch announcement; it should not be treated as an independent benchmark or proof of performance on every workload.
MiniMax M3.1 Flash features
Agentic tasks and workflow automation
MiniMax describes M3.1 Flash as better suited to complex workflows that span SaaS tools and multiple steps. Such work can include sales operations, marketing, customer service, finance, and HR processes. In practice, developers should check whether the model reliably selects the right tool, passes valid arguments, handles partial failures, and verifies the final state.
A useful agent evaluation measures completed tasks rather than counting tool calls alone. Record the percentage of tasks completed correctly, the number of retries, total tokens consumed, wall-clock time, and the need for human intervention. These measures help determine whether a claimed cost improvement translates into a lower cost per successful outcome.
Coding and software engineering
The launch also emphasizes software development. Coding agents must understand an existing codebase, identify relevant files, implement changes, run tests, and respond to failures. A model’s performance on a benchmark is useful context, but it cannot predict the result for every language, repository, framework, or development environment.
MiniMax cites its own evaluation results, including a 73.7% score on DeepSWE v1.1. That number is company-reported and should be interpreted in the context of the benchmark’s version, test setup, and scoring method. Teams considering M3.1 Flash for production coding should run it against representative tasks from their own repositories and compare the quality of the final patch, test pass rate, and cost.
For a broader workflow, compare it with the tools and models already used by your team. Articles About AI also covers practical developer infrastructure, including Amazon TorchNeuron and PyTorch on AWS Trainium, which explores a different part of the AI development stack: running workloads on alternative accelerator hardware.
Computer use
Computer-use models interact with graphical interfaces or software environments to carry out tasks. These systems can be useful when an application lacks a convenient API, but they introduce reliability challenges: layouts change, buttons move, sessions expire, and a seemingly successful click may not complete the intended operation.
MiniMax reports improvements in computer-use evaluations, including results from OSWorld V2 and Agents’ Last Exam. These figures are claims from the company announcement, not a guarantee that the model will reliably control every desktop or web application. Before deployment, test the exact browser, operating system, permissions, and workflow. Use explicit checkpoints for consequential actions and require confirmation before destructive or irreversible changes.
Cybersecurity tasks
MiniMax also highlights cybersecurity capabilities, including vulnerability discovery and validation. It reports an 82% result on CyberGym and describes internal evaluations involving code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering.
A benchmark score does not establish that a model is safe to use without supervision, nor does it mean that all security recommendations are correct. For defensive use, review findings with established tools and qualified security personnel. Restrict access to production systems, use isolated test environments, and validate suggested patches with tests and code review.
MiniMax M3.1 Flash API pricing
MiniMax publishes the following rates per one million tokens:
| Token category | Published price |
|---|---|
| Input tokens | $0.15 |
| Output tokens | $0.60 |
| Cached input tokens | $0.015 |
These are token prices, not a promise that every completed agent task will cost a fixed amount. A workflow can make multiple model calls, use long context, generate lengthy output, or repeat steps after an error. Consequently, the total cost depends on the complete task.
The company’s separate claim that agentic-task usage costs are 70% lower than its predecessor refers to overall usage cost for those tasks. It should not be confused with a 70% reduction in every token rate. To check the economics, run a representative batch of tasks on both models, record all token usage and retries, and divide total cost by the number of successfully completed tasks.
For an internal comparison, capture at least these metrics:
- Total input, output, and cached input tokens.
- Number of model calls and tool calls per task.
- End-to-end duration, including tool execution.
- Successful completion rate and test pass rate.
- Retry rate and human correction time.
- Cost per verified successful outcome.
That gives a more meaningful comparison than headline pricing alone.
Is MiniMax M3.1 Flash available now?
According to MiniMax’s October 19 announcement, M3.1 Flash is available through the MiniMax API, M Plan, and MiniMax Code. This is the company’s stated availability status at the time of writing. Actual access may depend on account type, region, subscription, and product-specific rollout conditions; developers should confirm access in the relevant official console before planning a deployment.
The model weights are not yet publicly released according to the same announcement. MiniMax says they will be released later, but the announcement does not establish a verified public download link or a confirmed release time. Do not assume that the model can already be self-hosted simply because the company has announced a future weights release.
Start with the official MiniMax M3.1 Flash announcement and the MiniMax API platform. Check the current model identifier, account eligibility, regional endpoint, rate limits, and terms before integrating it into an application.
How does the speed claim translate to real work?
MiniMax says online inference is nearly three times faster than its predecessor and that output speed can exceed 200 tokens per second in everyday use, with median P50 speeds above 170 tokens per second. These are vendor-reported measurements. The actual experience depends on context length, output length, server load, network latency, streaming configuration, and the number of tools in a workflow.
For a simple text response, faster token generation may reduce waiting time. For an agent, however, total completion time also includes planning, multiple model calls, external API latency, browser or computer actions, and verification. A faster model can still take longer to finish a task if it needs more retries or makes less accurate tool calls.
A fair benchmark should use the same prompts, tools, task set, and success criteria for both models. Run enough trials to account for variability and report median and percentile latency alongside task success. If the workload is coding, include test results and patch review; if it is computer use, verify the final application state rather than relying on the model’s description of what it did.
M3.1 Flash versus other coding and agent models
There is no universally best model for every coding or agent task. A useful comparison should include task success, reliability, latency, context handling, tool integration, price, and access requirements.
MiniMax’s published rates make M3.1 Flash worth testing for cost-sensitive workloads, especially where an application makes repeated model calls. However, token prices alone cannot establish that it is cheaper than a competing model for a particular task. Differences in output length, retries, context reuse, and tool orchestration can change the final cost.
Teams should create a small evaluation set from their real work: bug fixes, feature changes, test generation, documentation updates, and multi-step browser tasks. Use the same acceptance tests and human review rubric for each model. Keep model versions and configuration fixed during the comparison, and repeat the evaluation when providers change models or prices.
Limitations and questions developers should check
Several important limitations remain relevant:
- Vendor-reported performance: The headline speed, cost, and benchmark improvements come from MiniMax. Independent replication is needed to establish how well they generalize.
- No public weights yet: MiniMax says the weights will be released later, so self-hosting should not be assumed to be available today.
- Task-dependent costs: The 70% claim concerns overall usage costs for agentic tasks compared with the predecessor, not every possible workflow or token category.
- Tool reliability: Agent and computer-use tasks can fail because of external services, interface changes, permissions, or malformed tool calls.
- Regional and account access: The official product or API console is the right place to confirm access for a particular account and region.
- Security oversight: Cybersecurity benchmark results do not replace human review, sandboxing, or established security controls.
These are not reasons to dismiss the model. They are the checks needed to turn a launch announcement into a sound engineering decision.
A practical checklist before adopting MiniMax M3.1 Flash
- Confirm the current model ID, API endpoint, account eligibility, and pricing in the official documentation.
- Choose a representative set of coding or agent tasks rather than relying on generic demos.
- Define success criteria before testing, including correctness and the final state of external tools.
- Track tokens, retries, latency, tool costs, and human review time.
- Compare cost per successful task, not just cost per million tokens.
- Keep sensitive data out of prompts unless your organization’s data handling and contractual requirements permit it.
- Use least-privilege credentials and isolated environments for computer-use and security workflows.
- Re-test after model, API, pricing, or product-access changes.
What the launch means for AI developers
MiniMax M3.1 Flash adds another option to the market for coding and agentic AI. Its emphasis is clear: faster online inference and lower reported task costs while targeting software development, computer use, workflow automation, and cybersecurity. The published API rates are specific and useful, while the performance improvements should remain attributed to the company until independently tested.
For developers, the right next step is a controlled evaluation on real tasks. If M3.1 Flash maintains acceptable quality while reducing latency or the cost per completed workflow, it may be a strong candidate for selected workloads. If it requires more retries or human correction, the headline savings may not carry through to production.
Sources and further reading
- MiniMax official announcement: M3.1 Flash — Redefining Agent Performance and Cost — October 19, 2026.
- MiniMax API platform — official access point; verify current account eligibility and documentation.
- Reuters: China’s MiniMax launches updated AI model M3.1 Flash, claims major improvements — independent reporting on the announcement.