Quick Answer
MiniMax M3.1 Flash is a proprietary model for coding and AI-agent work that MiniMax says lowers usage costs by 70% and exceeds 200 tokens per second in typical generation. The model is available through MiniMax’s API, M Plan, and MiniMax Code. Compare the published token rates and plan limits before moving a production workload, because company-reported benchmarks do not replace independent testing.
Key Takeaways
- MiniMax released M3.1 Flash on October 19, 2026.
- The API starts at $0.15 per 1 million input tokens.
- MiniMax says the model can exceed 200 tokens per second.
- MiniMax Code supports macOS and Windows desktop users.
- Published benchmark scores are company-reported, not independent results.
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is a proprietary multimodal model aimed at AI agents, software development, and tasks that need a large working context. MiniMax rolled out the model on October 19, 2026, and said the anonymous “Space Bunny” system that processed more than 200 trillion tokens was an early version of M3.1 Flash. MiniMax’s launch announcement also says the model is available through the MiniMax API, M Plan subscription, and MiniMax Code.
MiniMax M3.1 Flash accepts text, image, and video inputs, according to an independent model catalog. The catalog lists a 1 million-token context window, which is the amount of information the model can consider within one request or conversation. Examenos’ model listing describes the model as proprietary and notes that MiniMax has not published total parameter counts or active parameter counts.
The large context window matters most for users who need to analyze long repositories, extensive documentation, or multi-step agent tasks without repeatedly splitting source material into smaller prompts. That said, context capacity does not prove that every answer will be accurate. Developers should still validate generated code, review tool actions, and keep a human approval step for consequential changes.
Where can users access MiniMax M3.1 Flash?
MiniMax M3.1 Flash is available through the MiniMax API, M Plan, and MiniMax Code. The API is the direct option for developers building the model into an application or automated workflow, while MiniMax Code is the company’s desktop coding environment. MiniMax says model weights will be released later, so users cannot currently treat the launch as a confirmed open-weight release.
MiniMax Code is offered as a desktop application for macOS and Windows. The app can connect to local files, tools, and a browser when the user authorizes those connections, according to the MiniMax Code download page. Local access can make an agent more useful for development work because it can use project files and development tools instead of relying only on copied text.
MiniMax Code’s local connections also create a practical permission decision. Review what folders, browser sessions, and tools the app can access before authorizing an agent, especially on a work computer or a device that contains private source code. Users who want to understand the broader tradeoff around AI systems that can act inside a desktop environment can also consider how local PC file access changes the data exposure question.
How much does MiniMax M3.1 Flash cost through the API?
MiniMax M3.1 Flash API list pricing is $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached-input tokens. MiniMax published those rates with the October 19 launch. Input tokens are the text or other material sent to the model, output tokens are the generated response, and cached-input tokens are reused prompt content priced at a lower rate.
MiniMax says M3.1 Flash lowers usage costs by 70% compared with its predecessor. That is a company claim rather than an independently verified cost comparison, and actual spending depends on prompt length, output length, caching behavior, retry rates, and how often an agent calls tools. The most useful comparison is the cost of completing a real workload, not only the listed price for a single token category.
| Cost category | MiniMax M3.1 Flash list price | Why it matters |
|---|---|---|
| Input tokens | $0.15 per 1 million tokens | Applies to prompts, instructions, and supplied context. |
| Output tokens | $0.60 per 1 million tokens | Applies to generated code, explanations, and agent responses. |
| Cached-input tokens | $0.015 per 1 million tokens | Can reduce costs when an application reuses the same context. |
MiniMax Code also lists three paid plans: Go at $22, Explore at $55, and Build at $132. The page displays promotional prices of $11, $27.50, and $66, respectively. Promotional pricing can change, so teams should confirm the current plan terms before assigning paid seats or relying on a discount for a long-term budget.
How fast does MiniMax M3.1 Flash generate responses?
MiniMax says M3.1 Flash can exceed 200 tokens per second in typical generation and has a median generation speed above 170 tokens per second. The company also says online inference is nearly three times faster than its predecessor. These are company-reported performance figures, so they describe MiniMax’s own measurement environment rather than a universal result for every prompt, region, or API workload.
Generation speed matters for agent workflows because a coding agent often performs several model calls before it completes a task. Faster individual responses can reduce the waiting time between planning, file inspection, code generation, and review. A fast model response does not eliminate other delays, including network latency, browser automation, tool execution, and the time required to run tests.
MiniMax M3.1 Flash may therefore feel most responsive when a task is limited to generation or analysis, rather than when an agent must operate across many external tools. Developers should test speed with their own request size, context length, and tool chain before treating a vendor performance claim as a production service-level expectation.
What do MiniMax’s benchmark results show?
MiniMax reports that M3.1 Flash scored 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, and 82% on CyberGym. These are company-reported benchmark results, not independent evaluations. The scores suggest that MiniMax is positioning the model for software engineering, computer-use tasks, broad agent evaluation, and cybersecurity-oriented testing.
Benchmark results are useful because they offer a structured way to compare a model’s performance on a defined test. However, the tests do not fully represent a company’s codebase, browser environment, internal security rules, or the quality of a developer’s prompts. A model can perform well on an evaluation and still make incorrect changes, misunderstand an undocumented requirement, or use an authorized tool in an unintended way.
The most sensible approach is to treat MiniMax’s published scores as a starting point for evaluation rather than a final purchasing decision. Run a limited trial on representative tasks, measure successful completion rates, inspect failures, and require code review before merging generated changes. Similar concerns apply when AI agents receive more autonomy, as shown by the questions raised around unauthorized agent actions in controlled testing.
What can MiniMax Code do on Windows and macOS?
MiniMax Code can run on Windows and macOS and can connect to local files, tools, and a browser after user authorization. Those connections allow a coding assistant to work with a local project rather than requiring a developer to paste every file into a chat window. The practical benefit is a more complete view of a project’s structure, dependencies, and related files.
MiniMax Code’s browser and tool access also mean users should treat authorization as a security boundary. Grant only the folders and tools needed for the task, avoid giving an agent access to credentials or unrelated personal files, and review proposed changes before allowing an automated workflow to proceed. Browser-connected AI features can be useful, but they also require clear limits on what data and actions an agent can reach.
MiniMax Code is best suited to developers who already understand their project environment and can recognize a risky or incorrect change. Users who are new to coding should begin with read-only tasks, such as explaining a file or suggesting a test plan, before allowing an agent to edit a working project. Stop using automated actions and contact the relevant platform or workplace administrator if an agent accesses data or makes changes outside the permissions you intended.
Who should consider MiniMax M3.1 Flash?
MiniMax M3.1 Flash is most relevant to developers and teams that need low listed API prices, long-context analysis, or a desktop coding tool with authorized local access. The model’s pricing is particularly notable for workloads with heavy prompt reuse because cached input costs less than standard input. Teams using long system prompts or repeated repository context may see more value from that pricing structure than teams with short, one-off prompts.
MiniMax M3.1 Flash is less suitable for organizations that require published parameter details, independently validated benchmarks, or immediate access to downloadable model weights. MiniMax has not published total or active parameter counts, and it says weights will arrive later. Those limits matter for buyers whose security review, deployment policy, or infrastructure plan depends on running a model in their own environment.
For most users, the practical next step is a controlled pilot. Start with a narrow coding or agent task, set clear file and tool permissions, compare the completed work against another model or manual process, and calculate full workflow costs. Developers evaluating AI tools should also keep security updates current because a compromised browser or operating system can expose the same files an authorized coding tool is designed to use.
FAQ
Is MiniMax M3.1 Flash available now?
Yes, MiniMax M3.1 Flash is available through the MiniMax API, M Plan, and MiniMax Code as of October 19, 2026. MiniMax says downloadable model weights will be released later.
How much does MiniMax M3.1 Flash cost?
MiniMax M3.1 Flash costs $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 per 1 million cached-input tokens at list price. Actual spending depends on prompt size, response length, cached context, and tool-related workflow calls.
Does MiniMax M3.1 Flash have a 1 million-token context window?
Yes, the Examenos model catalog lists MiniMax M3.1 Flash with a 1 million-token context window. A large context window can help with long documents and code repositories, but it does not guarantee accurate reasoning or correct outputs.
Can MiniMax Code access local files?
Yes, MiniMax Code can connect to local files, tools, and a browser after the user authorizes access. Review permissions carefully and avoid granting access to sensitive folders, credentials, or unrelated personal data.
Are MiniMax M3.1 Flash benchmark scores independently verified?
No, MiniMax’s published DeepSWE, OSWorld V2, Agents’ Last Exam, and CyberGym scores are company-reported results. Test the model on representative tasks before relying on those scores for a production decision.