MiniMax has launched M3.1 Flash, its latest lightweight frontier model, and confirmed that the anonymous model known as Space Bunny — which topped usage leaderboards for ten straight days while nobody knew who built it — was an early build of the new release. The Chinese lab announced the model on October 19, saying it cuts the cost of agentic work by about 70% against its predecessor while running online inference nearly three times faster.
The launch closes one of the odder episodes in the current model race: a model that proved itself in public, at enormous volume, before its maker put a name to it. MiniMax says Space Bunny processed more than 200 trillion tokens in total while operating anonymously. OpenRouter, which hosted the stealth run, gives a narrower figure for its own platform — 56.7 trillion tokens, the most any stealth model has processed there — and says the model held its number-one ranking for ten consecutive days before leaving the platform on October 5 and returning on October 19 as M3.1 Flash.
M3.1 Flash Pricing, Speed And Availability
MiniMax M3.1 Flash is priced flat at $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015 per million, according to the company’s launch post. The model carries a one-million-token context window and is available now through the MiniMax API, the company’s M Plan and its MiniMax Code desktop app. OpenRouter is offering it free for a limited introductory period.
On speed, MiniMax credits more aggressive sparsity in the architecture. It claims output can exceed 200 tokens per second, with median speed consistently above 170 tokens per second. OpenRouter’s own measurements from the stealth period point the same way: it says time to first token improved 38% during the anonymous run, throughput rose 45%, and the error rate fell from 1.4% to 0.7% as user feedback fed fixes back into the model.
M3.1 Flash Benchmarks: DeepSWE, OSWorld And The Hallucination Caveat
MiniMax positions the model for agentic tasks, coding, computer use and cybersecurity work. Its self-reported scores include 73.7% on DeepSWE v1.1, 61.7 points on OSWorld V2, 76.2 on BankerToolBench and 82% on CyberGym — figures the company presents as on par with top models at a fraction of the task cost, in the same cheap-and-fast tier where DeepSeek’s V4.1-Flash-1017 recently posted a leading DeepSWE score. As ever with vendor numbers, the benchmarks need reading as claims, not verdicts.
Independent measurement adds an important caveat. Artificial Analysis scores M3.1 Flash at 42 on its intelligence index — level with GLM-5.3-Flash and behind DeepSeek’s 45 — at roughly $0.22 per task. The same evaluation records a very high hallucination rate on its factual-accuracy test, because the model attempts nearly every question rather than declining when it does not know. That profile suits it to agent loops that verify their own work — the pattern behind releases such as Moonshot’s Kimi K3 — and makes it a riskier choice as a direct answer engine.
Weights, A Pro Model And What Comes Next
MiniMax says the model’s weights will be released later this month — reporting around the launch puts the date at October 23 — and that a Pro model, intended to push capability rather than price, will follow. Neither the weights nor the Pro model has a confirmed specification yet, and the company has not published a breakdown of the 200-trillion-token stealth figure beyond OpenRouter’s platform numbers.
The stealth launch itself is part of the story. Rather than announcing first and benchmarking later, MiniMax let an unnamed model absorb real developer traffic for weeks, fixed what broke, and only then attached its brand. It is a playbook other Chinese labs have edged towards, and it produced launch-day evidence — ten days at the top of a public usage chart — that no press release can manufacture.
Why It Matters
The central entity here is MiniMax M3.1 Flash, and its defining attributes are price, speed and proof of use. At $0.15/$0.60 per million tokens it resets the reference price for a competent agent-tier model, undercutting the cheap tiers of Western labs and competing directly with DeepSeek and Zhipu’s GLM line for developer traffic. Its neighbours are the routers and coding tools that distribute that traffic — OpenRouter, OpenCode, Cline and Ollama Cloud all carried the stealth run — and every lab whose “flash” model now has to justify a higher price against a rival that has already survived 56.7 trillion tokens of public use on one platform alone.
Background
MiniMax, founded in Shanghai in 2021, built its name on consumer apps such as Hailuo and Talkie before its M-series models made it a fixture of the open-weight race alongside DeepSeek, Alibaba’s Qwen and Moonshot’s Kimi. Its M3 model shipped with open weights in June 2026; M3.1 Flash is the efficiency play built on that base, aimed at the agent workloads where cost per task, not peak intelligence, decides which model gets the traffic.
Source: Unite.AI; Forkast — both reporting MiniMax’s official launch post and OpenRouter’s reveal, October 19, 2026.
0 Comments