MiniMax M3.1 Flash launches as the unmasked Space Bunny model with 70% cost cut

MiniMax has rolled out M3.1 Flash, the latest iteration of its fast model line, and used the launch to unmask one of the stranger experiments in recent AI releases: an anonymous model called Space Bunny was an early version of M3.1 Flash all along. According to MiniMax, Space Bunny handled more than 200 trillion tokens while operating under that name, and the launch announcement quickly reached Hacker News's front page.

MiniMax positions M3.1 Flash as a big step up for agentic work, coding, Computer Use and security tasks, while making online inference close to three times faster and cutting overall usage costs by 70% compared with its predecessor.

What the benchmarks show

All figures below come from MiniMax's own announcement rather than independent evaluation.

On agentic work, MiniMax says M3.1 Flash places among the strongest performers on AutomationBench-AA, a suite that tests whether a model can carry out genuine enterprise workflows spanning SaaS systems, covering areas such as sales, marketing, operations, customer service, finance and HR from start to finish. The model scores 76.2 on BankerToolBench, which MiniMax cites as evidence of strength in multi-step planning, keeping documents and files consistent, financial compliance and chained tool calls.

For software engineering, M3.1 Flash reaches 73.7% on DeepSWE v1.1, a benchmark built around complex tasks where a model has to understand a codebase on its own, build features, fix defects and verify its work through tests. MiniMax says the model holds up over long chains of steps, making it suited to full development workflows at lower cost.

In Computer Use, the category where models operate a computer and coordinate work across applications, the model scores 61.7 points on OSWorld V2 and 29.6 on Agents' Last Exam (ALE), a test focused on real professional work. MiniMax says it ranks first among Chinese models on both benchmarks.

On the security side, the model posts 82% on CyberGym, a benchmark for finding and validating vulnerabilities. MiniMax adds that its internal security team saw solid performance in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.

Speed and pricing

The efficiency gains come from what MiniMax describes as more aggressive sparsity in the model. In everyday use, output can exceed 200 tokens per second, with median (P50) throughput consistently above 170 tokens per second.

Pricing is flat per million tokens: $0.15 for input, $0.60 for output and $0.015 for cached input. MiniMax claims the overall cost of running agentic tasks is 70% lower than with the previous generation, and frames that combination, near-leading performance at a very low cost per task, as the model's core pitch across the AutomationBench, DeepSWE and ALE benchmarks.

Availability

M3.1 Flash is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Model weights will be published later, MiniMax says, and a separate Pro model intended to push the frontier of intelligence is also planned.

Why it matters

Two things stand out. First, the Space Bunny reveal is a data point on how large AI labs now test models in the open: by serving an unnamed version at scale, more than 200 trillion tokens is an enormous amount of real traffic, a lab can gather production signal on how a model behaves before attaching its brand to it. The tactic seems likely to spread.

Second, the economics. Agentic systems consume far more tokens than a single chat query, so a model that pairs frontier-adjacent benchmark results with a 70% drop in task costs changes the unit economics of building agents, particularly for high-volume automation across SaaS tools. The caveats are the usual ones: these are vendor-reported numbers on benchmarks the vendor helped popularise, and independent testing will show whether the claims hold. With open weights promised down the line, the wider community will eventually be able to check for itself.