The mask comes off

For 26 days, one of the most-used models on OpenRouter had no maker. Space Bunny Alpha appeared on September 23, free, with a million-token context and text, image and video input. Developers fingerprinted its tokenizer within days and pointed at MiniMax. MiniMax said nothing.

On October 19 the company ended the game in two steps. Its official account teased "Welcome back. Space Bunny" in the morning (09:21 UTC, per CellCog's tracking record), and its blog then rolled out MiniMax M3.1 Flash, stating plainly that Space Bunny "was an early version of MiniMax M3.1 Flash." Per CellCog's tracker, Space Bunny spent 25.8 days anonymous before the reveal.

An early version is not necessarily the released checkpoint. The usage, behaviour and results Space Bunny showed while anonymous do not automatically transfer to the model that went on sale today.

The model itself is not new in outline. A preview reached MiniMax Code and Token Plan subscribers on September 27 with no per-token price and no benchmark table. What changed today is everything a buyer actually needs: a public API, a rate card, scores, and a free listing on OpenRouter that developer reports spotted the same day.

Two token totals, neither scoped

MiniMax's blog says Space Bunny "processed more than 200 trillion tokens while operating anonymously." A separate OpenRouter usage summary, reported by TipRanks from the company's LinkedIn post, says Space Bunny held the top usage position for 10 days and processed 56.7 trillion tokens.

Those are two reported figures, not two measurements of the same thing. Neither published account states what its figure covers: which period, which routes, which model versions, or how tokens were counted. On the information available, the two numbers can neither be reconciled nor declared contradictory. The gap between them, roughly 3.5 to 1, is a comparison of reported figures, not a discrepancy in a common audited metric, and it would be premature to attribute the difference to any specific platform or route. Neither figure is a confirmed total across all platforms. Treat both as company claims until one of them is scoped.

The price is real. The 70% needs its workings

The rate card is the strongest part of the launch because anyone can verify it in a minute: $0.15 per million input tokens, $0.60 per million output, $0.015 per million cached input, flat. Against MiniMax's stated M3 list rates of $0.30 per million input tokens and $1.20 per million output tokens, the published rates are a halving. Actual savings for any given user still depend on provider pricing (third-party platforms list different rates for M3), cached-input usage and the workload.

MiniMax separately claims overall usage cost falls about 70% against the predecessor, and that inference runs nearly three times faster, with output reaching 200+ tokens per second and median speeds above 170. The public rate card alone does not reproduce the 70% figure. Verifying it would require the workload, the token mix, the caching assumptions and the billing methodology behind the company's calculation, and the blog shows none of that. The speed figures are likewise MiniMax's own, from its own serving stack.

The comparison that matters this week is not M3. DeepSeek shipped V4.1-Flash-1017 on October 17 at the same $0.15 / $0.60 headline, off-peak, per DeepSeek's official pricing page. DeepSeek doubles both prices at peak (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays) and charges $0.003 per million cached input tokens off-peak. At those cited off-peak cached-input rates, DeepSeek's $0.003 is one-fifth of MiniMax's $0.015. That comparison concerns cached input alone, not the total cost of running an agent, and it depends on DeepSeek's time-based pricing and cache eligibility. MiniMax charges one flat price and five times more for cached input. Which model is cheaper overall depends on when you run it and how much of your prompt repeats, and neither launch post will tell you that.

Scores, with the labels still on

Every benchmark in the launch post is MiniMax-reported: 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam and 82% on CyberGym. No independent evaluation had been published when this was written.

The automation score shows why the labels matter. MiniMax says M3.1 Flash ranks among the leading models on AutomationBench-AA, but the announcement's prose provides no numerical score. Numbers do circulate for similarly named automation evaluations, via a secondary OpenRouter summary and via readings of the blog's own charts, but they cannot be treated as interchangeable without confirmed benchmark versions and scoring conditions, so this article publishes no AutomationBench score. Until a third party runs the model, the honest summary is narrower: a fast, cheap agent model whose maker says it sits near the top on coding and computer use.

One more date to hold loosely. MiniMax's blog says weights "will be released later" and stops there. Reports describing a release later this week, and one specific date circulating in video coverage, do not come from the launch post. The free OpenRouter access is also reported rather than documented in the blog: TipRanks, summarising OpenRouter's LinkedIn post, describes a temporary 72-hour free-access window before paid pricing applies. Check the live OpenRouter listing for current availability and limits. A reported, time-limited promotion is not a price, and it does not make the regular MiniMax API free.

Sources

  1. [1] MiniMax, "MiniMax M3.1 Flash: Redefining Agent Performance and Cost" (official blog, October 19, 2026) — launch post: Space Bunny identity, pricing, benchmark table, speed and cost claims, weights "will be released later" with no dateRead source
  2. [2] OpenRouter, Space Bunny Alpha listing (revealed as MiniMax M3.1 Flash; 1M context; September 23, 2026) — checked live October 19, 2026Read source
  3. [3] TipRanks, "OpenRouter Showcases High-Usage MiniMax AI Model With New Pricing and Benchmarks" — OpenRouter summary via LinkedIn: 56.7T tokens, 10 days at top, 72-hour free windowRead source
  4. [4] CellCog, "Space Bunny Alpha Is MiniMax M3.1-Flash: The Record" — 09:21 UTC reveal post; 25.8 days anonymousRead source
  5. [5] DEV, "MiniMax M3.1 Flash just went free on OpenRouter" — free listing spotted October 19Read source
  6. [6] Frontier Now, "MiniMax M3.1-Flash: 1M context, subscription" — September 27 preview shape; M3 at $0.30/$1.20Read source
  7. [7] DeepSeek API Docs, Models & Pricing — official pricing page, checked October 19, 2026: deepseek-flash (V4.1-Flash-1017) at $0.15/$0.60 off-peak, peak rates double, cache hits $0.003 per 1M off-peak; peak hours 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidaysRead source