MiniMax confirmed on October 19, 2026, that a mysterious stealth model known only as Space Bunny was never a separate project at all. It was an earlier build of the company’s newly launched MiniMax M3.1 Flash, and it had quietly processed more than 200 trillion tokens while topping public leaderboards under an alias. The admission turns a months-long guessing game among AI developers into a confirmed data point, and it lands the same week rival labs are racing to ship cheaper, faster agentic models.
For anyone tracking the open leaderboard scene, the reveal answers a question that had been circulating since Space Bunny first appeared: who built it, and why did it vanish from view without ever being named. MiniMax’s answer, delivered alongside the official M3.1 Flash launch, folds that mystery directly into the company’s product story.
What Space Bunny Actually Was
MiniMax’s own account is specific: Space Bunny was an earlier version of MiniMax M3.1 Flash, and it had already “topped global leaderboards with 200T+ tokens” before the company put its name on it, according to the official announcement. That framing matters because it means the model developers had been quietly testing and benchmarking for weeks was not an experimental side project or a leaked competitor build. It was MiniMax’s own flagship in disguise.
Stealth testing of this kind is not unusual in the current AI model race. Labs frequently push unlabeled or pseudonymously named checkpoints onto public inference marketplaces to gather real-world usage data and leaderboard rankings before a formal launch, reducing the risk of a public benchmark miss. What is less common is a company explicitly confirming the identity of a stealth model after the fact, rather than letting the connection remain speculation. MiniMax did exactly that this week, tying Space Bunny’s token count and leaderboard standing directly to the production model it just released through the MiniMax API.
Several specific claims about Space Bunny’s run, including an exact streak length at the top of any particular leaderboard, a precise token count different from the 200 trillion figure MiniMax cited, a specific exit date, or a free-access promotional window, have circulated in community discussion but are not confirmed by MiniMax’s own statement. Readers should treat those extra details as unverified until MiniMax or the leaderboard operators themselves confirm them.
MiniMax M3.1 Flash: the Model Steps Out of Hiding
The production model now carries the official identifier MiniMax-M3.1-Flash and launched on October 19, 2026, through the MiniMax API, the company’s M Plan subscription tier, and its MiniMax Code offering. MiniMax positions the model as an agentic system built for tool use, coding, Computer Use tasks, cybersecurity workflows, and multimodal input, with a context window that stretches up to 1,000,000 tokens and a thinking-depth setting developers can tune for a given task.
That combination, an extended context window plus switchable reasoning depth, is the part of the release MiniMax is leaning on hardest. Rather than shipping one fixed model for every use case, the company is framing M3.1 Flash as adjustable: lighter reasoning for high-volume, low-latency agent calls, and deeper reasoning when a task actually needs it. For teams running thousands of automated agent calls per day, that flexibility can matter more than a single flashy benchmark number.
Benchmark Scores: Where M3.1 Flash Claims to Lead
MiniMax published four headline benchmark figures alongside the launch. On Automation-Bench, a measure of how well a model handles multi-step automated workflows, M3.1 Flash scored 62%. On ALE (Agents’ Last Exam), a tougher agentic reasoning test, it scored 30%. On DeepSWE, a software-engineering benchmark, it hit 74%. And on BabyVision, a visual-reasoning test, it reached 91%.
| Benchmark | What It Measures | MiniMax M3.1 Flash Score |
|---|---|---|
| Automation-Bench | Multi-step automated/agentic task completion | 62% |
| ALE (Agents’ Last Exam) | Advanced agentic reasoning under constraints | 30% |
| DeepSWE | Software-engineering and coding tasks | 74% |
| BabyVision | Visual reasoning and multimodal understanding | 91% |
The gap between the Automation-Bench score and the ALE score is worth noting on its own. A 62% result on a general automation benchmark next to a 30% result on a test built specifically to stress-test agentic reasoning suggests the model is more reliable at executing defined multi-step workflows than it is at open-ended, exam-style agentic problem solving. That is a useful distinction for any team deciding whether M3.1 Flash fits a structured automation pipeline versus a more exploratory agent task.
Pricing and the Million-Token Context Window
MiniMax M3.1 Flash is priced at $0.15 per million input tokens and $0.60 per million output tokens through the MiniMax API, with cached-input tokens billed at $0.015 per million, a 90% discount versus fresh input. Combined with the 1,000,000-token context window, that pricing structure is designed for workloads that repeatedly reuse large context blocks, such as an agent that keeps re-reading the same codebase, document set, or conversation history across many calls.
| Metric | MiniMax M3.1 Flash |
|---|---|
| Official model identifier | MiniMax-M3.1-Flash |
| Launch date | October 19, 2026 |
| Context window | Up to 1,000,000 tokens |
| Input pricing | $0.15 per 1M tokens |
| Output pricing | $0.60 per 1M tokens |
| Cached-input pricing | $0.015 per 1M tokens |
| Speed claim vs. M3 | Roughly 3x faster online inference |
| Cost claim vs. M3 | About 70% lower usage cost |
| Model weights release | October 23, 2026 |
A simplified example of how a developer might reference the model in an API call looks like this:
{
"model": "MiniMax-M3.1-Flash",
"messages": [
{"role": "user", "content": "Summarize this 500-page contract."}
],
"max_tokens": 4096
}That structure is illustrative of how the model identifier slots into a standard chat-completion style request; exact endpoint parameters should be confirmed against MiniMax’s own API documentation before production use. MiniMax also maintains a separate model list reference on its China-facing platform documentation, which developers outside the MiniMax API ecosystem can use to cross-check model identifiers.
Three Times the Speed, a Third of the Cost
MiniMax’s headline performance claim is that M3.1 Flash runs nearly three times faster in online inference than its predecessor while cutting usage costs by roughly 70%. A separate official MiniMax post phrases the same comparison slightly differently, describing the new model as running at three times the speed and one-third the cost of M3. Both framings point to the same underlying shift: MiniMax is optimizing this generation for throughput and cost efficiency rather than chasing a marginal accuracy gain.
That is a deliberate positioning choice. In a market where several Chinese AI labs are releasing updated models within weeks of each other, speed and price have become differentiators that developers can act on immediately, long before deeper accuracy comparisons settle. A model that is three times faster and 70% cheaper changes the cost calculus for any team running M3.1 Flash across thousands of daily agent calls, even before anyone fully evaluates whether its reasoning quality matches the prior generation.
Model Weights Arrive October 23
MiniMax’s initial announcement said weights for M3.1 Flash would be released at a later date without specifying when. A follow-up official MiniMax post then set that date as October 23, 2026, four days after the model went live through the hosted API. That gap between a hosted launch and an open-weight release has become a familiar pattern for MiniMax and several of its peers: ship the API first to capture usage and benchmark attention, then follow with weights once the hosted rollout has stabilized.
For the open-source AI community, the weights release matters more than the API launch in some respects. It is what allows independent researchers, smaller labs, and self-hosting developers to verify MiniMax’s benchmark claims directly, run the model on their own infrastructure, and build fine-tuned variants rather than depending entirely on MiniMax’s hosted pricing.
Why Labs Run Stealth Models on Public Leaderboards
The leaderboard incentive
Public inference marketplaces and leaderboards have become a proving ground that AI labs treat almost like a pre-launch focus group. By listing a model under an unrelated or invented name, a lab can collect real usage volume, measure how the model performs against live traffic rather than curated benchmark sets, and gauge standing against competitors, all without attaching its brand to a result that might embarrass it later. If the stealth model performs well, the lab can reveal the connection on its own timeline, turning months of anonymous leaderboard dominance into a ready-made launch narrative.
Why the reveal strategy is unusual
What makes MiniMax’s move notable is the explicit, named confirmation. Plenty of stealth models disappear from leaderboards without their creators ever acknowledging the connection, leaving the identification to community speculation. MiniMax instead built the Space Bunny reveal directly into its M3.1 Flash announcement, turning 200 trillion tokens of anonymous leaderboard activity into evidence of the production model’s real-world readiness on day one.
From M1 to M3.1 Flash: MiniMax’s Model Line
MiniMax M3.1 Flash is the latest step in a model lineage that has moved quickly over the past year, from the company’s earlier M1 and M2 generations through the M3 release and now to this faster, cheaper M3.1 Flash variant. Each generation has pushed MiniMax further into agentic and coding-oriented use cases rather than general chat, a strategy that puts it in direct competition with other fast-moving Chinese labs chasing the same agentic-AI market. Readers can find tech-insider.org’s coverage of the original M3.1 Flash launch specs for a closer look at the model’s initial benchmark and pricing rollout, and the earlier three-way comparison against GPT-6 Luna and DeepSeek V4.1 Flash for how it stacked up against rivals before this week’s reveal.
The pace of that lineage, several named generations inside roughly a year, mirrors a broader trend across Chinese AI labs in 2026: shipping cheaper, faster, more narrowly agentic models on a near-monthly cadence rather than waiting for a single large leap. MiniMax’s willingness to run a flagship anonymously first, then reveal it as a finished product, fits that same compressed release rhythm.
Competitive Landscape: DeepSeek, GLM, Kimi and the Race for Cheap Agentic AI
MiniMax M3.1 Flash arrives into a crowded field of low-cost, high-context Chinese models built for agentic and coding work. DeepSeek updated its own V4.1-Flash-1017 model under an MIT license just days before MiniMax’s announcement, and the company separately opened a search API alongside a DeepSWE score of 77.0, putting pressure on MiniMax’s own DeepSWE claim of 74%. Reports have framed the two launches as part of the same competitive wave, with MiniMax, DeepSeek, GLM, and Kimi all shipping flash-tier, agent-focused models within a matter of weeks of each other.
That overlap puts buyers in an unusually favorable position. When three or four labs are releasing comparable context windows, comparable tool-use capabilities, and comparable per-token pricing within the same month, the practical differentiator for most teams becomes reliability and latency under real production load rather than a half-point benchmark gap. It also means MiniMax’s Space Bunny reveal functions as a marketing tool in a race where attention is scarce: a 200-trillion-token anonymous track record gives the company a concrete, verifiable talking point that a benchmark chart alone does not provide.
The coding-benchmark race in particular has become a proxy war between these labs. Tools built on top of these models, including coding assistants and editor integrations, are increasingly judged on which backend model developers choose by default, a dynamic that has already played out in comparisons like JetBrains’ own Mellum 2.1 coding model, which climbed to a 47% SWE-Bench score this quarter.
Market Impact: What the Reveal Means for AI Buyers
For enterprise buyers evaluating agentic AI vendors, the Space Bunny reveal is a reminder that leaderboard rankings on open inference marketplaces are not always what they appear to be. A model sitting near the top of a public leaderboard under an unfamiliar name may well be a major lab’s unreleased flagship rather than an independent newcomer, which changes how much weight that ranking should carry in a procurement decision. Teams that track leaderboard standings as part of their vendor evaluation process now have fresh reason to ask who actually built a high-ranking anonymous model before assuming it reflects a smaller, less-resourced competitor.
There is also a direct cost angle. With input pricing at $0.15 per million tokens and output at $0.60 per million, M3.1 Flash lands in the same aggressive pricing tier that has made Chinese-built models attractive to cost-sensitive developers throughout 2026, a trend that has already pushed government and enterprise policy discussions, including scrutiny visible in the broader debate over DeepSeek’s access to US federal IT systems. Pricing this low on a model with a million-token context window puts additional pressure on US-based providers to justify premium pricing with measurable reliability or compliance advantages rather than raw capability alone.
Predictions: Where This Goes Next
- Expect at least one more lab to confirm a previously anonymous leaderboard entry as its own flagship within the next quarter, following MiniMax’s lead in turning stealth-testing data into launch marketing.
- The October 23 weights release will likely trigger a wave of independent benchmark re-runs from the open-source community, some of which may not match MiniMax’s self-reported Automation-Bench and ALE figures exactly.
- Pricing pressure from M3.1 Flash and DeepSeek’s V4.1-Flash-1017 will likely push at least one additional competitor, plausibly GLM or Kimi, to announce a comparable price cut before the end of 2026.
- Enterprise procurement teams will increasingly request disclosure clauses asking AI vendors whether any of their public leaderboard entries are undisclosed internal models, a direct response to episodes like this one.
- MiniMax Code and the M Plan subscription tier are likely to see the fastest early adoption of M3.1 Flash, since both channels were built specifically around agentic and coding workloads rather than general consumer chat.
How M3.1 Flash Compares on Paper
Taken together, the benchmark table and the pricing table above paint a model built for volume rather than peak accuracy on the hardest agentic tests. A 30% score on ALE is a reminder that even a well-funded, fast-moving lab like MiniMax has not solved the hardest class of agentic reasoning problems, even as its 91% BabyVision score and 74% DeepSWE score show real strength in visual reasoning and software tasks. That split is consistent with how MiniMax has pitched the model: not as a single best-in-class reasoner, but as a fast, cheap, flexible workhorse for the agentic and coding tasks that make up the bulk of real production traffic.
Developers deciding between MiniMax M3.1 Flash and competing flash-tier models should weigh the million-token context window and sub-$1-per-million-token blended pricing against the specific benchmark profile their workload actually needs, rather than defaulting to whichever model posted the single highest headline number this week.
Frequently Asked Questions
What is Space Bunny?
Space Bunny was a stealth AI model that appeared anonymously on public leaderboards and processed more than 200 trillion tokens. MiniMax confirmed on October 19, 2026, that it was an earlier version of its new MiniMax M3.1 Flash model.
What is MiniMax M3.1 Flash?
MiniMax M3.1 Flash, officially identified as MiniMax-M3.1-Flash, is an agentic AI model from MiniMax built for tool use, coding, Computer Use, cybersecurity tasks, and multimodal input, with a context window of up to 1,000,000 tokens.
How much does MiniMax M3.1 Flash cost?
Through the MiniMax API, input tokens cost $0.15 per million, output tokens cost $0.60 per million, and cached-input tokens cost $0.015 per million.
How much faster is M3.1 Flash than the previous MiniMax model?
MiniMax claims M3.1 Flash runs roughly three times faster in online inference than M3, at about 70% lower usage cost, with one official post describing it as three times the speed and one-third the cost of M3.
When will MiniMax release the model weights for M3.1 Flash?
MiniMax’s initial announcement said weights would come later without a fixed date. A subsequent official MiniMax post set the release date as October 23, 2026.
What benchmark scores did MiniMax M3.1 Flash report?
MiniMax reported 62% on Automation-Bench, 30% on ALE (Agents’ Last Exam), 74% on DeepSWE, and 91% on BabyVision.
Where can developers access MiniMax M3.1 Flash?
The model is available through the MiniMax API, the company’s M Plan subscription, and MiniMax Code.
Is the claim that Space Bunny topped OpenRouter for exactly 10 consecutive days confirmed?
No. MiniMax’s official statement confirms Space Bunny processed more than 200 trillion tokens and topped global leaderboards, but a specific 10-day streak, an exact token figure different from that 200 trillion mark, a precise exit date, or a free-access promotional window are not confirmed by MiniMax’s own announcement and should be treated as unverified.