Evidence at a glance
An Available API, Not Yet an Open Technical Design
MiniMax released M3.1 Flash on October 19, 2026, making it available first through its API, M Plan, and MiniMax Code. It targets coding, office agents, and visual reasoning. The multimodal mixture-of-experts model has 428 billion total parameters, about 23 billion active parameters, and a one-million-token context window. It accepts text, image, and video inputs and returns text; the product page also mentions sparse attention and a native visual encoder.
For technical leaders, the important question is not simply that the model is large but activates relatively few parameters. Those specifications do not explain what changed in the Flash variant. Its total and active parameter counts match MiniMax M3, and MiniMax has not published a technical report for M3.1 Flash. Architecture labels alone cannot establish inference efficiency, training changes, or reliability over long contexts, and a million-token window does not guarantee dependable performance on million-token tasks.
Its Best Results Are Concentrated in Specific Agent Tasks
All currently available scores are reported by MiniMax, with no independent replication yet. M3.1 Flash scores 62.2 on Automation-Bench v1.0.6, ahead of Opus 5 at 50.3 and GPT-5.6 Sol at 45.8 in MiniMax's table. It ties Opus 5 at 76.2 on BankerToolBench and scores 60.3 on LOCA-Bench 256K. This supports a limited conclusion: the model is competitive on several office and tool-use tasks, not that it has proven broadly superior agent capabilities.
Its coding results resist being reduced to a single verdict. On DeepSWE v1.1, it scores 73.7, slightly above GPT-5.6 Sol at 73.0 but below DeepSeek-V4.1-Flash-1017 at 77.0. Its Terminal Bench 4.0 score is 38.1, compared with 51.8 for Opus 5, and its SWE Refactor Bench score of 23.0 trails Opus 5 at 31.0. For vision, it scores 91.0 on BabyVision with tools and 82.5 on CharXiv RQ. These scores should be interpreted within their individual tasks, not treated as direct estimates of production success.
The Cost Question Is Per Task, Not Just Per Token
The API is priced at $0.15 per million input tokens and $0.60 per million output tokens, with cached reads at $0.015. MiniMax puts greater emphasis on cost per task: its Automation-Bench chart says all five reasoning levels cost less than $0.05 per task, while scores rise from 52.8 at Low to 62.8 at Max. Another point of reference is Agents’ Last Exam, where a 152-task run cost about $111 and scored 29.6. These figures help with budgeting, but they do not establish whether the answers meet a business quality threshold.
There is also a discrepancy in the release materials that should remain visible: the Automation-Bench chart gives Max a score of 62.8, while the score table lists 62.2. Those values should not be silently collapsed into one definitive number. The materials also say that on DeepSWE, DeepSeek V4.1 Flash-1017 scores higher at its top setting while costing less. In other words, M3.1 Flash is not simply “cheap” in the abstract; cost depends on the task, reasoning level, token volume, and comparison model.
Always-On Reasoning Changes the Integration
Reasoning is always enabled in M3.1 Flash and cannot be switched off as it can in some models. Developers can choose among Low, Medium, High, Xhigh, and Max, with MiniMax recommending Max. Requests that set thinking.type=disabled or effort=none return an error. For existing M3 integrations, this is more than a renamed parameter; the calling contract has changed.
Reasoning is returned in a separate reasoning_content field, which may be absent. Code that previously parsed inline reasoning tags from the response needs to change, and clients must reserve output budget for both reasoning and the final answer. In an agent pipeline, this affects parsing, context management, and token budgeting. Before deployment, teams should test retries, missing fields, and truncation on long tasks rather than merely swapping the model name.
Validate the Workload Before Waiting for the Weights
The model is available now through the MiniMax API, with Anthropic-compatible and OpenAI-compatible access, including Chat Completions and Responses. Weights are expected on October 23, 2026, but the materials do not disclose the hardware required for self-hosting. MiniMax says an earlier stealth build called Space Bunny processed more than 200 trillion tokens. That suggests substantial prior use, but it is not a substitute for external benchmark replication and does not answer questions about deployment cost or self-hosting requirements.
A more defensible decision is to run a small comparison on representative in-house tasks. Start with office agents or tool use, where the reported results are strongest, and include harder terminal work in the style of Terminal Bench. Track success rates, human rework, total token cost, and failure modes. If a critical workflow depends on disabling reasoning, or most of its workload is demanding coding where results are weaker, a low input price may not offset quality and integration costs. Reassess self-hosting after the weights arrive rather than committing resources before the hardware requirements are known.