MiniMax M3.1 Flash Slashes AI Agent Costs 70%

Chinese AI developer MiniMax on October 19 officially released MiniMax M3.1 Flash, a faster and cheaper update to the M3 large language model it launched in June, aimed squarely at the ballooning cost of running AI agents. According to the company, agentic tasks cost 70% less to run on MiniMax M3.1 Flash, while inference runs nearly three times faster, topping 200 tokens per second in everyday use.

The launch takes aim at a well-known pain point in the agent economy: a single agentic task can require dozens of model calls, turning token pricing into the dominant line item for anyone running autonomous workflows at scale. MiniMax framed MiniMax M3.1 Flash as its "ongoing exploration of the balance between efficiency and capability," and priced it aggressively — a flat $0.15 per million input tokens puts it head-to-head with rival Chinese labs in a growing price war. SiliconANGLE reports that the flat rate matches the off-peak price of DeepSeek's V4.1-Flash, a competing model upgraded just days earlier.

What MiniMax M3.1 Flash is

MiniMax M3.1 Flash is a multimodal mixture-of-experts model listed at 428 billion total parameters with roughly 23 billion active, according to MarkTechPost, with MiniMax crediting the extra speed to more aggressive sparsity at inference time. The model accepts text, image and video input, and prompts can run to 1 million tokens of context — enough to feed entire codebases or long document sets into a single request.

One design choice will matter to developers migrating from M3: reasoning stays on for every request in MiniMax M3.1 Flash and cannot be disabled. Developers pick one of five effort levels — Low, Medium, High, XHigh and Max — with MiniMax recommending Max as the default. Requests that attempt to switch reasoning off return an error, and the model's reasoning arrives in a separate output field reserved for it, so integrations that parsed inline thinking tags from M3 will need updating before they can call the new model.

Benchmarks and the Space Bunny stealth run

In MiniMax's own testing, MiniMax M3.1 Flash scored 61.7 on OSWorld V2, which measures how well a model can operate a computer — the company said no Chinese model has done better. It posted 76.2 on BankerToolBench, a banking workflow test, and 82% on CyberGym, the UC Berkeley-built benchmark for reproducing real software vulnerabilities, with MiniMax listing code auditing, penetration testing and malware analysis among the jobs the model can take on. On the coding benchmark DeepSWE v1.1, MiniMax M3.1 Flash managed 73.7%, just behind DeepSeek V4.1-Flash's 77% — results from both labs that have not been independently checked.

The most unusual part of the launch is the model's backstory. MiniMax confirmed that "Space Bunny," an anonymous model that had been running in stealth on OpenRouter, was an early build of MiniMax M3.1 Flash. MiniMax claims the stealth build processed more than 200 trillion tokens in total; OpenRouter, which hosted the run, counts 56.7 trillion tokens on its platform and says Space Bunny was its number-one model for 10 straight days. Unite.AI reports that OpenRouter credits user feedback from the stealth period with concrete fixes — time to first token improved 38%, from 1.95 seconds to 1.22 seconds, throughput rose 45%, and the error rate fell from 1.4% to 0.7%. The stealth model left OpenRouter on October 5 and returned October 19 as MiniMax M3.1 Flash, which OpenRouter is offering free for 72 hours.

The pricing war — and what it means for agent builders

The headline number for agent builders is the cost per task. For MiniMax M3.1 Flash, the company charges $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens; because the model uses fewer output tokens per task, the effective cost works out to about $0.22 per task, according to Oossa. That is half the standard rate of the bigger M3, and — notably — half of what DeepSeek charges for V4.1-Flash, which MiniMax's rivals updated on October 17.

The aggressive pricing lands in a week already crowded with agent-infrastructure announcements, from Together AI's post-training service to new enterprise cloud offerings like Vultr's agentic AI platform upgrade. Cheap, fast frontier-adjacent models are the substrate the whole agent economy runs on, and MiniMax is betting that shaving 70% off agent run costs will buy MiniMax M3.1 Flash developer loyalty at exactly the moment agent frameworks are standardizing on low-cost reasoning models.

MiniMax M3.1 Flash is live now on the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Open weights are slated for October 23, according to the company's post on X, and MiniMax says a Pro version will follow at an unspecified date. Independent benchmarks and real-world agent deployments over the coming weeks will show whether the company's "generational leap" claims hold up outside its own test suite — but for teams watching their agent token bills, the math behind MiniMax M3.1 Flash is already hard to ignore.