TL;DR
MiniMax launched M3.1 Flash on October 19 through its API, M Plan, and MiniMax Code. The company claims nearly 3x faster inference and 70% lower agent-task usage cost than its predecessor, and says the anonymous Space Bunny was an early version that processed more than 200 trillion tokens. Its benchmarks, speed, cost, and usage figures have not been independently validated for this briefing.
What MiniMax announced
On Monday, October 19, Hong Kong-listed MiniMax rolled out MiniMax M3.1 Flash, its next-generation Flash-tier model. The company says it delivers a generational leap in four areas: complex agentic tasks, coding, computer use, and cybersecurity. Online inference is nearly three times faster, and the overall usage cost for agentic tasks is 70% lower than its predecessor, according to the company.
MiniMax lists prices of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. It reports output speeds of 200+ tokens per second in everyday use, with median (P50) speeds of 170+ TPS, attributing the efficiency to more aggressive sparsity in inference. These are vendor-reported performance figures; the listed token rates do not guarantee a particular saving per completed task.
M3.1 Flash is available now through the MiniMax API, M Plan subscription, and MiniMax Code desktop app. MiniMax says model weights and a Pro variant will arrive later; the announcement gives no release dates or sequence.
Space Bunny unmasked
MiniMax identified the anonymous model codenamed Space Bunny as an early version of M3.1 Flash. The company says it processed more than 200 trillion tokens while operating anonymously. This briefing has not independently verified that usage total.
The useful next question is how the named, priced service performs on builders' own workloads. Its reported stealth usage is a separate claim from measured task quality, latency, or cost.
The benchmarks — all company-reported
On agentic work, MiniMax says M3.1 Flash ranks among the top performers on AutomationBench-AA and scored 76.2 on BankerToolBench, which tests multi-step planning, financial compliance, and chained tool calls. On coding, it reports 73.7% on DeepSWE v1.1. These are MiniMax-reported results; comparisons with other vendors require checking evaluation settings and reproducing the tests.
On computer use, MiniMax reports 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam, and says those results rank first among Chinese models on both. On cybersecurity, it claims 82% on CyberGym for vulnerability discovery and validation. Every figure and ranking in this section comes from MiniMax's announcement and has not been independently validated for this briefing.
What matters for builders
Our view: the listed $0.15/$0.60 input/output token rates make M3.1 Flash worth comparing on agent workloads. The advertised 200+ TPS and 70% task-cost reduction remain MiniMax claims, so builders should measure completed-task cost, quality, and latency under their own conditions. Future weights could enable self-hosted evaluation, but local-hardware suitability cannot be assessed until model size, license terms, and deployment requirements are known.
Independent testing can start now through the available API; a weights release is not required to test the hosted service. Future weights would add the option to test self-hosted deployments. Benchmark reproduction and corroboration of MiniMax's claimed 200-trillion-token stealth usage remain separate questions.