Tool Suggestions
MiniMax M3.1 Flash: 3x Faster, 70% Cheaper — Built for Agentic Coding
On October 19, MiniMax released M3.1 Flash, a multimodal model built for coding agents, automated office work and computer use. The pitch is straightforward: near-frontier performance on agentic tasks, at a flat price of $0.15 per million input tokens and $0.60 per million output tokens. MiniMax says the model runs at roughly three times the speed of its M3 predecessor and cuts the cost of agentic tasks by about 70%. Both figures are the company's own claims, and they are the reason this release matters: agent workloads consume tokens in volume, so a cheaper, faster model changes what teams can afford to automate. M3.1 Flash is also the release version of "Space Bunny," a stealth model that MiniMax says processed more than 200 trillion tokens while running anonymously on OpenRouter.
What the model is
M3.1 Flash is a mixture-of-experts model with 428 billion total parameters and roughly 23 billion active per token, according to MiniMax's product page. It uses sparse attention and a native visual encoder, accepts text, image and video input, and returns text. The context window is 1,000,000 tokens, which MiniMax's documentation states covers input and output combined. Those headline numbers match the earlier M3, and MiniMax has not published a technical report explaining what changed inside the Flash variant. One design choice does stand out: reasoning is always on and cannot be disabled. Developers pick from five effort levels — Low, Medium, High, Xhigh and Max — and MiniMax recommends Max as the default. Requests that try to switch thinking off will return an error, which is a real consideration for anyone migrating an existing integration.
The numbers MiniMax is claiming
The pricing is the clearest part of the announcement. The rate card is flat: $0.15 for input, $0.60 for output and $0.015 for cached input, per million tokens, with no surcharge for long contexts. The older M3 costs $0.30 input and $1.20 output for requests up to 512K input tokens, and its rates double above that threshold. Per token, then, Flash is half of M3's standard rate. The widely quoted "70% cheaper" figure is different: MiniMax says it reflects the total cost of completing agentic tasks in its own evaluations, helped by the model needing fewer tokens per task. It is a claim about task economics, not a tariff cut, and MiniMax has not published the calculation behind it.
On speed, MiniMax credits "more aggressive sparsity" and claims output can exceed 200 tokens per second in everyday use, with a median (P50) consistently above 170 tokens per second — nearly three times the online inference speed of M3. The company's product page lists a more conservative figure of about 150 tokens per second. On benchmarks, all reported by MiniMax, M3.1 Flash scores 73.7% on DeepSWE v1.1 for software engineering, 62.2 on Automation-Bench v1.0.6 for multi-step workflow automation, 76.2 on BankerToolBench, 61.7 on OSWorld V2 for computer use, 29.6 on Agents' Last Exam and 82% on the CyberGym security benchmark. These are vendor-run results; no independent verification was available at the time of writing.
Who it is for, and how to try it
The intended user is a team running agents at volume: coding assistants that work through repositories, automation pipelines that call tools in long sequences, or subagents handling routine steps under a larger model. At these prices, cost per completed task matters more than any single benchmark, and MiniMax's own cost chart places M3.1 Flash at roughly a quarter of a dollar per DeepSWE task, against several dollars for frontier models. The model is live now on MiniMax's platform as MiniMax-M3.1-Flash, with both OpenAI-compatible and Anthropic-compatible API access. It is also available through the M Plan subscription, the MiniMax Code desktop app and OpenRouter. Model weights have not been released yet; MiniMax's launch post pointed to October 23, and the license terms have not been disclosed, so self-hosting is not an option today.
The honest limits
Three cautions apply. First, every performance figure here comes from MiniMax itself, and its own comparison table shows the model is not the leader everywhere: on Terminal Bench 4.0 it scores 38.1 against 51.8 for Claude Opus 5, and DeepSeek's V4.1 Flash posts a higher DeepSWE score at a similar price point. Second, the always-on reasoning and the new reasoning_content response field mean integrations built for M3 — including any that parsed inline think tags — need updating before they work. Third, with no technical report and no weights yet, claims about what improved over M3 rest entirely on the company's word. None of this cancels the appeal of the pricing, but it does mean the sensible move is a small pilot on your own workload before committing an agent pipeline to it.
The takeaway
M3.1 Flash is a serious attempt to make agentic coding and automation cheap enough to run at scale: a 1M-token context, flat $0.15/$0.60 pricing and vendor-claimed speeds above 170 tokens per second at the median. The 3x speed and 70% cost figures should be read as MiniMax's claims, not established facts, and the benchmark lead is real only on some tests. For teams whose agent bills are the bottleneck, it earns a trial. For everyone else, it is worth watching until the weights and independent results arrive.