MiniMax releases M3.1 Flash with $0.15/M input pricing and 1M context

MiniMax says the model has about 23 billion active parameters, accepts text, image and video input and returns text, and targets coding, agentic office work and visual reasoning. Its product page lists sparse attention and a native visual encoder. The parameter totals match MiniMax M3, and MiniMax has not published a technical report on what changed in the Flash variant. MiniMax confirmed that Space Bunny, a stealth model that processed more than 200 trillion tokens, was an earlier M3.1 Flash build.

Reasoning is always on and cannot be disabled. Developers can choose five effort levels—Low, Medium, High, Xhigh and Max—and MiniMax recommends Max as the default. Requests that send thinking.type=disabled or effort=none return an error, and reasoning arrives in a separate reasoning_content field that may be absent; integrations that parse inline tags from M3 need updating. The model is live on platform.minimax.io as MiniMax-M3.1-Flash and is available through Anthropic-compatible and OpenAI-compatible access, including Chat Completions and Responses, as well as M Plan and the MiniMax Code desktop app. MiniMax says weights are due Oct. 23, 2026, but has not disclosed the license or self-hosting hardware.

All benchmark scores cited by MiniMax are vendor-reported and have no independent replication, the report said. On MiniMax’s table, M3.1 Flash leads Automation-Bench v1.0.6 at 62.2, ahead of Opus 5 at 50.3 and GPT-5.6 Sol at 45.8; it ties Opus 5 on BankerToolBench at 76.2 and tops LOCA-Bench 256K at 60.3. Agents’ Last Exam lands at 29.6.

Coding results are mixed. DeepSWE v1.1 scores 73.7, just above GPT-5.6 Sol at 73.0, but DeepSeek-V4.1-Flash-1017 scores higher at 77.0. Terminal Bench 4.0 is the weakest result at 38.1, versus 51.8 for Opus 5, and SWE Refactor Bench sits at 23.0, versus 31.0 for Opus 5. On vision, BabyVision with tools reaches 91.0 and CharXiv RQ reaches 82.5; HLE without tools scores 43.4.

MiniMax frames value as cost per task rather than per token. On its Automation-Bench chart, all five effort levels sit below $0.05 per task, with scores ranging from 52.8 at low effort to 62.8 at max. A full 152-task Agents’ Last Exam run cost about $111 at 29.6. MiniMax’s cost chart plots Automation-Bench max effort at 62.8, while its table says 62.2, an inconsistency in its materials.

API pricing is a flat $0.15 input and $0.60 output per 1 million tokens; cache reads cost $0.015 per 1 million, with no long-context surcharge. MiniMax M3 costs $0.30/$1.20 up to 512K input and doubles above that. MiniMax says Flash costs one-third of M3 but has not detailed the calculation; per token, Flash is half of M3’s standard rate.

MiniMax’s product page lists about 150 tokens per second, and the announcement claims three times the speed of M3. In MiniMax’s comparison table, M3.1 Flash’s 73.7 DeepSWE score trails DeepSeek V4.1 Flash’s 77.0, while its 62.2 Automation-Bench score is above DeepSeek’s 54.8, Kimi K3’s 46.7 and GLM-5.3’s 48.8; DeepSeek V4.1 Flash plots both higher and cheaper at max effort on DeepSWE. For scale, M3’s MXFP8 build is about 440 GB, but Flash’s self-hosting hardware was not disclosed.

The report cautioned that every benchmark is vendor-reported and that harness choice shifts results. Automation-Bench numbers do not transfer across sources: Kimi K3’s model card reports 30.8 on a 600-task public subset, while MiniMax lists 46.7. Spot checks hold up for some figures, with Kimi K3’s card reporting DeepSWE 67.5 and BabyVision 85.7, matching MiniMax’s table.