MiniMax unveils M3.1 Flash with 70% lower agent costs

If you noticed a mystery model named Space Bunny quietly chewing through over 200 trillion tokens on OpenRouter recently, the disguise is finally off.

MiniMax officially rolled out M3.1 Flash. The new model is the full, production-ready version of that stealth project, aimed directly at agentic workflows, software development, computer use, and cybersecurity.

Instead of chasing raw size, MiniMax optimized for speed and cost. Inference speeds are nearly three times faster, while the overall cost for running agent tasks drops by 70%.

Serious throughput on a budget

Anyone building autonomous agents knows the pain: once a model starts chaining tool calls, latency spikes and the token bill explodes.

M3.1 Flash tackles this by leaning heavily into aggressive sparsity. Third-party trackers note it keeps a 428B total parameter architecture with only 23B active parameters, remaining much lighter than rivals like DeepSeek Flash at 763B.

That lean footprint pays off in raw speed. Normal output regularly tops 200 tokens per second, with median P50 speeds hovering around 170+ TPS. Responses arrive almost instantly.

Pricing is equally aggressive: flat rates are $0.15 per million input tokens, $0.60 for output, and just $0.015 for cached inputs. For multi-step agent workloads, that translates to a 70% discount compared to previous setups.

Built to control screens and write code

The real focus here is getting work done in real environments.

On DeepSWE v1.1, a benchmark for autonomous software engineering, M3.1 Flash scored 73.7%, proving it can parse existing codebases, build features, and verify fixes without hand-holding.

It also took the top spot among Chinese models on two major agent benchmarks: 61.7 on OSWorld V2 for cross-application computer operation, and 29.6 on Agents' Last Exam (ALE) for complex professional labor.

Elsewhere, it notched 76.2 on BankerToolBench for multi-step compliance workflows and 82% on CyberGym for vulnerability discovery.

Access is already live via the MiniMax API, M Plan subscriptions, and the MiniMax Code desktop app. The lab noted that model weights will be released publicly at a later date, alongside a more capable Pro tier.

Why it matters

MiniMax is making a very smart, very practical bet here. Instead of fighting vanity wars on generic chatbots, it is turning its Flash tier into an ultra-cheap, high-speed execution engine for agents.

For developers and startups, this is an immediate win. Running computer-use loops and continuous coding agents usually burns through budgets in hours. At fifteen cents per million input tokens and near-instant token generation, autonomous workflows finally make economic sense.

For the broader industry, it proves that smaller active parameter counts like 23B can deliver top-tier agent performance when paired with sharp sparsity. It puts real pressure on competing open-weight labs to justify their heavy inference overhead.

The catch is patience. While API and app access are live right now, self-hosters will have to wait for the open-weight release. And because this is the efficiency-first Flash model, those needing peak frontier reasoning will likely need to hold out for the upcoming Pro version.