On the evening of October 19, MiniMax officially closed the biggest guessing game of the past two weeks: Space Bunny, the zero-priced anonymous model that had been absorbing real traffic on OpenRouter and OpenCode since September 23, is MiniMax M3.1 Flash. The announcement arrived with benchmark scores, API pricing, and integration details, plus confirmation that the model weights will be released later. The headline numbers first: more than 200T tokens processed cumulatively during the anonymous period (official figures), a 70% drop in comprehensive usage cost on agent-class tasks versus the previous generation, and online output speeds above 200 tok/s. A model with no known provenance sat a 13-day exam graded by real traffic and only then signed its name. The playbook deserves as much attention as the scorecard, and this piece takes both apart.
Scope note: this article draws on reporting by Suanlibao, DOIT, PANews, and Tencent News from around October 20, 2026, cross-checked against MiniMax's official announcement. All benchmark scores below are vendor-reported (official figures) with no independent reproduction yet. Prices are snapshots; the official pricing page prevails. Not investment advice.
1. A 13-Day Timeline: The Full Life of a Zero-Priced Space Bunny
Start with the sequence. On September 23, Space Bunny Alpha went live on both OpenRouter and OpenCode: no vendor attribution, no model card, and an input/output price of exactly zero. Its only self-introduction was the spec sheet itself — a 1M context window, text/image/video multimodal input, tool calling, and five reasoning-effort tiers. Free, anonymous, and impressive on paper, it quickly accumulated a first wave of genuine users.
The community began fingerprinting within a day. On September 24, developers compared tokenizers and token counts and found a strong match with the MiniMax family; around the same time, others noticed that test code in the MiniMax Code repository had prematurely included a MiniMax-M3.1 configuration. The company never confirmed or denied, and the speculation kept compounding.
On September 27, the MiniMax Code platform quietly listed an M3.1-Flash-Preview — no announcement, no benchmarks, a textbook case of doing without telling. On September 30, Space Bunny recorded the highest single-day call volume of the month on OpenRouter; third-party tracking put its peak-week token consumption above 28T. Keep these figures distinct: the single-day record is platform-verifiable, the 28T figure comes from third-party tracking, and the officially cited 200T+ cumulative total covers the entire anonymous window. They live on different time axes and should never be blended into one claim.
On October 1, a banner appeared announcing that service would end on October 5 — the first model in the stealth anonymous series to publish its own shutdown date in advance; previous anonymous animals came and went without notice. On October 5, Space Bunny went offline as scheduled. On the evening of October 19, MiniMax claimed it officially. On weight release, the official phrasing is "to be open-sourced later" with no exact date, while multiple reports suggest October 23 — a detail covered further below.
2. The Anonymous Grey-Release Playbook: What a Vendor Buys with 13 Days of Free Compute
Space Bunny was not the first anonymous animal, but it was the loudest. The playbook has three precedents: Pony Alpha was unmasked as Zhipu's GLM-5, Hunter Alpha as Xiaomi's MiMo, and Ox Alpha as GLM-5.3 Flash. Compared with those, Space Bunny absorbed the most traffic and ran the longest anonymous window — 13 days, over twice the shortest precedent.
The design deserves a close reading because it is, in essence, a controlled experiment. Anonymity plus zero pricing means every single call is made purely on model experience: no brand halo, no price anchor, no fandom filter. What the vendor collects is the cleanest evaluation data available — the composition of real workloads, the distribution of task types, the volatility of peak hours, and, most precious of all, long-tail failure modes. Internal benchmarks can never manufacture these, because internal evaluation never meets real users at their most adversarial. Publishing benchmarks after the trial inverts the usual order — promise first, verify later becomes verify first, promise later — and while the risk is front-loaded onto the vendor, credibility compounds.
Why did MiniMax need this? The backdrop is the chilly reception of M3: ninth among mainstream models on Artificial Analysis's intelligence index, a rating downgrade from JPMorgan, and a market capitalization that has shrunk more than 80% from its peak. In that environment, any high-profile launch would have been read through an existing negative narrative. The anonymous grey-release sidesteps the baggage: the model banks a 200T-token report card on experience alone, and the news that MiniMax shipped a new model lands only after users have already tried it and liked it. For a company near a trust low, this is textbook public-relations engineering.
The community's 13 days of field testing deserve an honest record, both sides. On the positive side, several developers observed self-verification behavior: the model would screenshot intermediate results to check its own work mid-task, and fix problems it had noticed before declaring completion — rare in this price class. On the negative side, hallucination was repeatedly flagged as high; consistency across runs with the same prompt was poor, raising stability concerns; and the default reasoning tier is max, so even trivial tasks burn heavy reasoning tokens. That was harmless while free, but a real line item once paid launch began. Yet set against the product's five-tier reasoning design, they inadvertently validate why the tiers exist.
3. The Scorecard and the Price Tag: Official Numbers, Aligned Before Judged
On weights: the official wording is that they will be open-sourced "later," and multiple reports point to October 23 — treat the date as unconfirmed until the checkpoint actually lands.
The October 19 announcement included the following benchmark results (all vendor-reported official figures, no independent reproduction yet):
| Benchmark | M3.1 Flash result | Note |
|---|---|---|
| DeepSWE v1.1 | 73.7% | Coding agent |
| OSWorld V2 | 61.7 | Officially claimed best in China |
| Agents' Last Exam | 29.6 | Officially claimed best in China |
| CyberGym | 82.2 | Security tasks |
| Automation Bench v1.0.6 | 62.8 | Some media cite 62.2 |
| BankerToolBench | 76.2 | Financial tool calling |
| LOCA 256K | 60.3 | Long context |
| Terminal Bench 4.0 | 38.1 | Terminal tasks |
| HLE | 43.4 | General reasoning |
Two details must be flagged. First, the "best in China" claim officially covers only OSWorld V2 and Agents' Last Exam — two benchmarks, and nothing beyond them. Second, Automation Bench circulated in the media with two values, 62.2 and 62.8; the difference traces to two scoring approaches, the table figure versus a re-test at the max reasoning tier. The 62.8 is the official table number, the 62.2 comes from the other measurement, and both are given here for readers to weigh. The official materials also offer a striking derived figure: passing all 152 questions of Agents' Last Exam is estimated to cost about 739 yuan (official estimate, referencing third-party reproductions of the predecessor).
On efficiency and cost, the official figures read: online inference speed up nearly 3x versus the previous generation; output speed above 200 tok/s (P50 above 170); and a 70% reduction in comprehensive usage cost on agent-class tasks.
Pricing drew the most media attention, and it exists in two sets of figures that should be presented side by side. The English-language report (PANews) gives $0.15 per million input tokens and $0.60 per million output tokens. The Chinese-language report (Tencent News) gives 1 yuan input, 4 yuan output, and 0.1 yuan cache-read per million tokens. Converted at exchange rates the two sets do not reconcile exactly, and the tier definitions may differ; the company has not mapped one onto the other. Both are reported here without adjudication — for procurement, the official pricing page is the authority.
One easily missed detail directly affects buying decisions: the thinking process cannot be turned off. Attempting to disable reasoning through the API returns an error; only the reasoning-effort tier is adjustable. For latency-sensitive, budget-sensitive, or determinism-oriented workloads, this means even trivial tasks incur reasoning overhead, and that line belongs in your cost model before you integrate.
Consider what the spec combination means in practice. A 1M context window plus native video input gives M3.1 Flash a distinct surface on agent workloads: a full screen recording, a complete session of operation traces, or dozens of pages of product documentation can all go into a single request. For automation scenarios that need to watch the screen before acting, the traditional pipeline slices the video, extracts frames, and runs OCR and captioning separately before feeding text to a model; when one model takes the raw input natively, both the pipeline cost and the error propagation across stages disappear. That is also the premise of its lead on cross-SaaS automation benchmarks — though it remains an inference from vendor-reported numbers until third-party evaluations weigh in.
4. The Architecture Base: MSA Sparse Attention Carried Over, No Tech Report for the Flash Changes
M3.1 Flash inherits the MoE specification of M3: 428 billion total parameters with 23 billion activated, built on MiniMax's proprietary MSA sparse attention. The official description is a two-stage design: a lightweight index first filters top-k KV blocks, then the main branch computes precisely over the selected blocks. Per-token compute is about one-twentieth of the predecessor's, with prefill accelerated more than 9x and decoding more than 15x. That architecture is the technical origin of the 70% cost reduction on agent tasks — long-context, multi-turn tool-calling agent workloads are exactly where sparse attention pays off most.
Honesty requires the caveat: the specific changes that make the Flash variant different from the standard M3.1 have no technical report behind them yet, and this article will not speculate. On open-sourcing, the official wording is "later," with no confirmed date, while multiple reports point to October 23. License terms, commercial-use provisions, and other key details should be confirmed against the official documentation when the weights actually land — by industry convention, all of that remains subject to change until then.
5. The Flash Twins: One Pushes Code to the Limit, the Other Bets on End-to-End Business Loops
Placed in its price class, the most natural yardstick for M3.1 Flash is DeepSeek-V4.1-Flash-1017. The official-table comparison:
| Benchmark (official figures) | MiniMax M3.1 Flash | DeepSeek-V4.1-Flash-1017 |
|---|---|---|
| Terminal Bench 4.0 | 38.1 | 42.3 |
| DeepSWE v1.1 | 73.7% | 77.0% |
On coding-oriented benchmarks DeepSeek is stronger — it leads on both Terminal Bench 4.0 and DeepSWE — and its weights ship under the MIT license and can be self-hosted, which is better for engineering autonomy. MiniMax's edge sits on the other side: it leads on cross-SaaS automation and Computer Use tasks, and it is the only model in this class supporting video input. Browser operation, interface understanding, multimodal agents — the end-to-end business-loop scenarios are its home ground.
The route split is clean: DeepSeek pushes engineering code to its limit; MiniMax bets on end-to-end business automation. For readers, this is not a choice between good and bad but a question of workload profile: writing code, refactoring repositories, driving terminals — the DeepSeek line fits. Operating software, running processes across systems, understanding interfaces and video — the MiniMax line fits. This site has a dedicated piece unpacking the DeepSeek-V4.1-Flash-1017 weight update; read together, the two cover both routes in this price band.
One practical suggestion: inventory your workloads before choosing. If 80% of a team's daily volume is code review, repository refactoring, and terminal work, the model with the stronger coding benchmarks is the lower-risk pick. If the core workflow moves data between web-based SaaS tools — taking screenshots, clicking buttons, filling forms — then the differentiation of native video input and Computer Use is worth more than a few points on any single benchmark. And if budget allows, running both models for a week on your own real tasks beats every scorecard; the 62.2-versus-62.8 discrepancy on Automation Bench alone shows how wide the variance between measurement approaches can be.
6. The Compliance Footnote: The Free Tier Has a Price Buried in the Terms
Having praised the paradigm's value, its other side must be stated plainly. The anonymous-period terms of service for Space Bunny stated that user prompts and completions may be retained. The anonymous service carried no SLA, no failure fallback, and no support channel — if something broke, there was nowhere to appeal. And the parties responsible for data ownership and security were ambiguous throughout the anonymous window: you did not know exactly whom you were signing a de facto data agreement with.
Over 13 days, 200T tokens of real traffic flowed in, which means a large number of companies and developers — knowingly or not — sent real business data into an unsigned black box. A hobbyist running toy tasks on a free model has little at stake, but if your calls carried customer data, internal code, or sensitive material, the anonymity was financed by opaque data flows. This playbook has now been validated as a traffic detonator and will almost certainly be reused, so this footnote is written not only for this particular space bunny but for the next one.
One-line close: When 200T tokens of anonymous traffic speak before the benchmark table does, and the unmasking lands with input pricing at the 1-yuan level, the try-first-review-later launch model has run its first full loop — but in the free controlled experiment, nobody stands guard over the participants' data flows.
Sources
- MiniMax official announcement (2026-10-19): M3.1 Flash launch and Space Bunny attribution
- Suanlibao (2026-10-20): M3.1 Flash launch coverage
- DOIT (2026-10-20): M3.1 Flash launch coverage
- PANews (2026-10-20): pricing and benchmark coverage of M3.1 Flash
- Tencent News (2026-10-20): Chinese-language pricing coverage of M3.1 Flash
Related reading
- DeepSeek Weights Are Free, Yet Agent Task Bills Fall 60 Percent: DeepSeek Weights Are Free, Yet Agent Task Bills Fall 60 Percent
- MiniMax Opens Its Deck: mcode, the Terminal Agent You Can Audit: MiniMax Code CLI: M-series models in your terminal
- Seedance 2.5 vs MiniMax H3 Review: 30-Second Narrative vs Audio Multi-Shot: Seedance 2.5 vs MiniMax H3: video generation compared