MiniMax M3 API: Pricing, Playground and Examples

Project coding: fix and iterate on existing code

We would prioritize MiniMax M3 for clearly defined iterations, such as fixing existing issues, adding features according to interface contracts, or adjusting implementations based on test results. Thinkwright's public engineering experiment found that M3 and GLM-5.2 performed similarly on fixes and feature additions to existing code, with the more noticeable gap concentrated in building projects from scratch.

Based on this, we suggest trying M3 for ongoing improvements and new feature iterations in existing projects.

Long-document analysis: read codebases, documents, and runtime logs together

MiniMax M3 uses MSA sparse attention to handle long context. Its technical paper describes selecting relevant information blocks before computing attention, reducing the computational overhead of very long inputs.

We focus more on its value for analysis across materials, such as checking whether an implementation matches requirements, comparing interface contracts across versions, or reconstructing an incident from logs. We recommend retaining file names, versions, and timestamps, and requiring conclusions to reference specific locations. For very long tasks, the model can first list the key evidence it found before making judgments.

Image and video understanding: include interfaces and visual materials in analysis

We would use MiniMax M3's visual capabilities for work requiring visual understanding, such as checking whether a webpage screenshot matches an implementation, analyzing chart contents, or documenting a workflow from an operation video. Like the official MiniMax API, API Hop provides image and video inputs, using image_url content blocks for images and video_url content blocks for video.

We recommend making your requirements explicit: identify the areas to focus on when checking an interface, specify the metrics to verify when analyzing charts, and limit the time range and events to identify when processing video. This makes it easier to align model output with the original materials and continue changing code or adding explanations.

Tool collaboration: continue multi-step tasks using intermediate results

MiniMax M3 supports tool calling and can continue reasoning across multiple rounds of tool interaction. Once an application provides retrieval, code execution, or file-processing tools, the model can decide what to do next based on their results. The official tool tutorial also requires retaining the complete model response in subsequent turns to maintain continuity of task information.

Customer feedback on M3's agent experience is mixed: some users appreciate its execution capabilities, while others report confusion in complex workflows. We therefore recommend starting with short, clearly bounded workflows, specifying each step's input, expected result, and failure handling before gradually extending task length.

Long-running optimization: keep improving against verifiable goals

MiniMax demonstrated an FP8 matrix-multiplication kernel optimization case in which M3 made 147 benchmark submissions and 1,959 tool calls over approximately 24 hours, continuously revising its implementation based on feedback.

We believe this capability is worth applying to work with clear goals, such as code performance optimization, test fixes, and experiment reproduction. We recommend defining correctness checks, performance targets, and stopping conditions in advance; when results stop improving for an extended period, check the task description, tool feedback, or implementation direction rather than merely increasing the number of calls.

Software engineering: 59.0% on SWE-bench Pro and a SWE-fficiency score of 34.8

In the official release results, MiniMax M3 scored 59.0% on SWE-bench Pro and 34.8 on SWE-fficiency. We use these as references for project issue fixes and code performance optimization, respectively.

SWE-fficiency focuses on speed improvements while preserving functional correctness, with metrics involving the relative performance of model and expert optimizations. We retain 34.8 as the benchmark score to avoid misreading it as a 34.8% speedup in a real project. For performance optimization tasks, we recommend checking runtime, resource usage, and test results together.

Terminal execution: 66.0% on Terminal-Bench 2.1

MiniMax reported an M3 score of 66.0%. The test used Terminus 2, a two-hour timeout per task, and a maximum output budget of 128K tokens. These evaluation settings help explain the execution conditions behind the score.

We focus more on whether the model can understand feedback after executing commands and continue making corrections. For practical evaluation, we recommend tasks that require installing dependencies, running tests, and handling errors, retaining logs from failed steps to check whether the task is ultimately completed.

Tools and retrieval: 74.2% on MCP Atlas and approximately 83.5% on BrowseComp

In the official report, MiniMax M3 scored 74.2% on MCP Atlas and approximately 83.5% on BrowseComp. We use these results as references for tool collaboration and complex retrieval capabilities.

For workflows that search multiple materials and verify conditions before producing conclusions, we recommend checking "finding relevant content" and "answering the question correctly" separately. For example, require the model to retain source links, explain which conditions the evidence supports, and flag unresolved questions. We prioritize checking whether information returned by tools actually supports the final answer.

Multimodal understanding: 91.6 on OmniDocBench and 78.1% on MMMU-Pro

In the complete official evaluation table, MiniMax M3 scored 91.6 on OmniDocBench using the v1.5 dataset and 78.1% on MMMU-Pro. We believe these results are useful for examining document parsing and multidisciplinary visual understanding, respectively.

OmniDocBench evaluates document elements including text, tables, and formulas. We would prioritize testing M3 on materials with complex layouts, nested tables, or formulas, while checking that extracted content corresponds to its page location. For questions requiring further reasoning, we recommend also verifying the data used in that reasoning, rather than only checking whether text recognition is complete.

Independent evaluation and response speed: an index score of 29 and output at approximately 100 tokens/second

Artificial Analysis scored the thinking version of MiniMax M3 at 29 on Intelligence Index v4.3.2; output speed through the MiniMax API was approximately 100.4 tokens per second, with a time to first token of approximately 2.65 seconds.

Compared with MiniMax M2.7: larger gains in terminal and tool tasks

In the official release table, MiniMax M2.7 and M3 scored 51.1% and 66.0% on Terminal-Bench 2.1, and 49.4% and 74.2% on MCP Atlas, respectively.

For teams already using M2.7 and planning to upgrade to M3, we recommend prioritizing multi-turn tool calls, code fixes, and long conversations in replacement evaluations. For short tasks that existing workflows already complete reliably, we would continue comparing response quality, duration, and call costs to assess the actual benefit of upgrading.

Compared with Claude Opus 4.7: higher retrieval scores, but a software engineering gap remains

In the same release table, M3 and Claude Opus 4.7 scored approximately 83.5% and 79.3% on BrowseComp, compared with 59.0% and 64.3% on SWE-bench Pro. We therefore consider M3's retrieval performance noteworthy, while software engineering capabilities still need task-specific comparisons.

We recommend testing information retrieval, cross-source verification, and project fixes separately. If the work relies mainly on retrieval, prioritize evaluating M3; if it requires completing a substantial change in a complex codebase in one attempt, we would retain Opus 4.7 as a comparator and observe pass rates and subsequent rework.

Compared with GPT-5.5: similar code-fix scores, with a larger terminal-execution gap

In MiniMax's release table, M3 and GPT-5.5 scored 59.0% and 58.6% on SWE-bench Pro, compared with 66.0% and 78.2% on Terminal-Bench 2.1. GPT-5.5's terminal score was cited from the benchmark's official leaderboard, and its test environment was not fully aligned with M3's. Official MiniMax M3 release and evaluation notes

We interpret the first pair as similar performance on that evaluation; a difference of 0.4 percentage points alone does not support a claim of a clear lead. For tasks requiring successive commands, environment debugging, and completed delivery, we recommend further comparing final completion rates rather than choosing a model solely from its code-fix score.

Compared with GLM-5.2: a public engineering experiment shows a trade-off between speed and complete delivery

In the Thinkwright experiment published by Brandon Huey on June 19, 2026, 60 scored tasks were run three times per model. The experiment used Fireworks priority service with thinking disabled on the client. The results were as follows. Thinkwright first-hand engineering experiment

We interpret these results as a completion-rate versus waiting-time trade-off within that experiment: M3 was faster, while GLM-5.2 had a higher full pass rate. The larger gaps recorded by the author were concentrated in build-from-scratch tasks, including package structure and interface behavior. We therefore recommend prioritizing clearly bounded modification tasks for M3 and adding checks for entry points, directory structure, and boundary conditions when delivering new projects.

Usage-based billing and model cost optimization

Charges follow current API Hop rates and actual usage. Top up on demand.

Unified AI models and data APIs, with multi-model access in one place

One API key connects multiple AI models and data services, reducing integration costs and letting you switch with one click. Business applications need both AI and data.

Call logs and application access controls, with clarity on every call

Every API call is recorded in your dashboard. Create separate API keys for different applications.

Evaluate outputs against your task

Use the documented model ID and supported parameters through API Hop. Evaluate outputs against your own acceptance criteria, then inspect usage records and response metadata before production use. A model name or example does not guarantee each result.

Prepare your account and test online

Sign in, confirm your balance, and create an API key.

Try it in the Playground

Explore the model's capabilities in the Playground to find the best way to use it.

Integrate MiniMax M3 into your business

Follow our developer documentation to integrate MiniMax M3 into your business.

Future model upgrades

To switch to a newer model, first confirm that its exact ID is returned by the model catalog for your API key. If it supports the same endpoint and request fields, update the model ID, then test response formats, streaming, output limits and billing before sending production traffic. For asynchronous image or video tasks, also verify job submission, status polling and result retrieval.