MiniMax M3.1 Flash Is the Viral ‘Space Bunny’ Model — and It Promises 70% Cheaper Agents

Yes: Space Bunny was MiniMax M3.1 Flash all along. On October 19, the Chinese lab confirmed that the anonymous model that quietly topped OpenRouter's charts for ten straight days — chewing through 56.7 trillion tokens on that platform alone — was an early build of its new agent model. And the pitch is aggressive: 70% lower agent costs than its predecessor, roughly 3x faster inference, priced at $0.15 per million input tokens and $0.60 per million output tokens.

This is the stealth-to-scale playbook, and MiniMax just ran it better than anyone has before. Instead of a flashy keynote, the company dropped an unnamed model into the wild in late September, let developers beat on it for weeks, and only attached a name once the thing had already won. OpenRouter says Space Bunny hit number one in daily token volume within three days and stayed there, processing more tokens than any stealth model in the platform's history. MiniMax itself puts the total stealth traffic above 200 trillion tokens.

The stealth run wasn't just marketing — it was free QA. OpenRouter told MiniMax exactly what broke, and the company says it shipped fixes before launch: time to first token improved 38% (from 1.95 seconds to 1.22), throughput rose 45% (70 to 100 tokens per second), and the error rate halved from 1.4% to 0.7%. That's the real story here. A ten-day, 56-trillion-token public stress test tells you more about a model's readiness than any benchmark table.

MiniMax M3.1 Flash specs and pricing

The model is a mixture-of-experts design with 428 billion total parameters and roughly 23 billion active per token — the same rough shape as the flagship M3, tuned for speed rather than maximum capability. Here's what matters:

• Context: 1 million tokens in, up to 524K tokens out
• Inputs: text, image, and video; text output only
• Built for: agentic workflows, coding, computer use, cybersecurity
• Pricing: $0.15 per million input tokens, $0.60 per million output tokens, $0.015 for cached input — about half of M3's rates
• Reasoning: always on, cannot be disabled. You pick the effort level (low, medium, high, xhigh, max); it defaults to max, and trying to switch thinking off returns an error
• Availability: MiniMax API, M Plan subscription, and MiniMax Code now; open weights promised October 23; a larger Pro model is teased
• Free window: live on OpenRouter at $0 for 72 hours starting October 19

The always-on reasoning is the detail developers will feel most. No toggle, no escape hatch — if your integration parses M3's inline think tags, it needs updating. The upside is a dial instead of a switch: run quick edits at low effort and nasty refactors at max without swapping models. The downside is that you pay for thinking on every call, and the default is the most expensive setting.

MiniMax M3.1 Flash benchmarks — all company-reported, none independently verified

Now the part that needs a big red flag. Every benchmark number below comes from MiniMax's own materials. No third party has independently verified them, and the company hasn't published a technical report or model card. Treat this section as the company's pitch, not settled fact.

With that said, here's what MiniMax claims: 62.2 on Automation-Bench (topping its own comparison table, and in its own graph beating Anthropic's Opus 5.5 at roughly one-twentieth the cost), 61.7 on OSWorld V2, 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, and 82% on CyberGym. The weak spots are in the table too: 38.1 on Terminal Bench 4.0 against 51.8 for Opus 5, and mid-pack scores on hard coding.

Context cuts both ways. The stealth traffic is genuine third-party evidence that the model holds up under real load — 200 trillion tokens don't lie about stability. But stability isn't capability, and MiniMax's own documentation has small inconsistencies (one analysis found the Automation-Bench score listed as 62.2 in the table but 62.8 in the graph). Meanwhile DeepSeek V4.1 Flash, released two days earlier, beat M3.1 Flash on the DeepSWE coding test. The honest read: strong on agentic office work and automation, ordinary on hard coding — until independent testers weigh in.

The quiet reroute nobody's code asked for

Launch day came with a footnote that deserves more attention than the benchmarks. Starting October 19, all M2.5 and M2.7 subscription calls automatically route to M3.1 Flash. The old model names keep working, your allowances don't change, and nothing returns an error — but the model answering is one nobody chose. Both old families shut down entirely on November 18.

This is worth a hard look because of the economics. MiniMax is a listed company and its API business is reportedly around 63% of revenue. M3.1 Flash bills at half of M3's standard rate, its reasoning can't be switched off, the default effort is max, and thinking tokens draw from the same allowance — while per-minute request ceilings drop from 500 to 200. From the user's side, you get a faster, probably better model. From MiniMax's side, every subscriber's consumption just went up on a cheaper per-token price. Whether that math works for the company is the number to watch, not the benchmark table.

What to do this week

The smart move is boring: try it while it's free. The 72-hour OpenRouter window is the cheapest model evaluation you'll ever run — point your existing agent stack at minimax/minimax-m3.1-flash:free, run your actual workload, and see if the cost-per-task beats what you're paying now. If you're weighing it against the field, our agent wars rundown has the broader context on where the cheap-agent tier is heading.

Watch October 23 for the open weights, which would let anyone self-host and settle the pricing question for good. And keep one habit: whenever a lab's launch slide shows a score-per-dollar graph with its own model in the top-left corner, wait for someone without a financial interest to reproduce it. MiniMax earned its 200 trillion tokens of proof. It hasn't yet earned the right to skip the independent test.