OpenRouter 'Space Bunny Alpha' Concludes Stealth Testing: MiniMax M3.1 Identity Reveal Imminent

The high-profile anonymous preview model 'Space Bunny Alpha' (stealth/space-bunny-alpha) on global AI routing gateway OpenRouter has reached its scheduled stealth testing expiration date of October 5, 2026. With the catalog listing's designated preview window elapsing, developer and community trackers anticipate an imminent reveal of the underlying foundation model.

Initially listed on OpenRouter under the anonymous provider moniker 'Stealth' on September 23, 2026, Space Bunny Alpha quickly climbed to the platform's top tier by token volume due to its zero-cost ($0 in / $0 out) availability. Coupling a massive 1,000,000-token context window with native multimodal capabilities and an un-bypassable reasoning architecture, the model generated widespread discussion across the developer community, prompting deep forensic analysis into its origins.

1M Context Window and 524K Completion Ceiling: The Unprecedented Specifications

Space Bunny Alpha's meteoric adoption was driven primarily by an architectural envelope that went well beyond standard flash-tier offerings.

According to OpenRouter's model catalog and API endpoint payload, Space Bunny Alpha offered a rare configuration:

• 1,000,000-Token Context Window: An expansive 1M input window enabled developers to ingest entire production repositories, image collections, or video files for whole-context synthesis.
• 524,288-Token Maximum Output Ceiling: Across all 466 models available on OpenRouter, exactly one model reported a maximum completion limit of 524,288 tokens: Space Bunny Alpha. This ceiling significantly outstripped typical commercial models (which cap output between 4,096 and 16,384 tokens) and exceeded the 512,000-token cap of MiniMax's public M3.
• Native Multimodal Input: The model natively supported text, image, and video inputs, synthesizing visual and textual data into reasoned text outputs.
• Mandatory 5-Tier Reasoning Architecture: The reasoning subsystem could not be disabled; submitting requests with effort: "none" triggered an HTTP 400 error requiring adaptive thinking. Users could select between low, medium, high, xhigh, and max, with the endpoint defaulting to max when omitted—delivering deep chain-of-thought exploration across complex engineering workflows.

Identical Tokenizers and 5-Point Parameter Fingerprint: Why Evidence Points to MiniMax M3.1

Although OpenRouter's listing described the provider simply as a third party that "chosen to remain anonymous during this preview," independent researchers began dissecting the model's behavioral fingerprint within hours of launch.

The most concrete clues emerged from independent tokenizer matching across several evaluation labs:

• Third-Party Tokenizer Verification: Following deployment, independent evaluator cheaty demonstrated that Space Bunny Alpha's text and vision tokenizers matched MiniMax M3 with 100% precision. MarMar Labs subsequently verified that 36 input-token comparisons across 12 distinct text fixtures aligned identically with MiniMax tokenizers, while cleanly diverging from Qwen, GLM, DeepSeek, and Llama families. YFarmX echoed these findings, verifying an exact match across 50 separate benchmark test strings.
• Five-Point Convergence with MiniMax Documentation: The strongest direct evidence surfaced on MiniMax's developer documentation (platform.minimax.io). MiniMax had quietly documented an unannounced internal model, MiniMax-M3.1-Flash-Preview, gated behind its enterprise "M Plan" and "MiniMax Code" offerings. Its documented schema aligned with Space Bunny Alpha across five independent criteria:
  1. Exactly 1,000,000 tokens of context
  2. Mandatory reasoning (effort: "none" rejected)
  3. The identical five-rung effort ladder (low to max)
  4. Defaulting to max when unspecified (unique among OpenRouter catalog entries)
  5. Native text, image, and video input support

This five-way parameter convergence across 466 catalog candidates established an estimated 85% confidence consensus among researchers that Space Bunny Alpha is an early deployment of MiniMax's upcoming M3.1 series.

OpenRouter's 'Alpha' Stealth Lineage and What Comes Next

Space Bunny Alpha represents the fourth iteration in OpenRouter's lineage of anonymous 'Alpha' preview releases.

Previous stealth entries on the platform similarly served as unannounced live testbeds for prominent AI labs before adopting their official commercial identities:

• Hunter Alpha was later unmasked as Xiaomi's MiMo-V2.5.
• OX Alpha became Z.ai's GLM 5.3 Flash.
• Union Alpha transitioned to Unbiased Pareto.

As the October 5 expiration window elapsed, OpenRouter prepared to de-list the anonymous preview route and transition users to the production endpoint under its formal laboratory branding.

While the forensic evidence linking the model to MiniMax M3.1-Flash is comprehensive, industry observers note that official laboratory confirmation remains pending until OpenRouter and MiniMax publish joint release statements. Developers should also anticipate that the promotional zero-cost pricing structure and the extraordinary 524K token output ceiling may transition to standard commercial tiered rates and production rate limits once the official model goes live.

Sources

• YulmuCoffee (@yulmu_coffee): Space Bunny Alpha Identity Reveal Lead
• HunterAlphaHub: Space Bunny Alpha: Specs, Identity Lead and the 5 October Exit
• CellCog: Space Bunny Alpha: The Stealth Model and the MiniMax Clues
• Codersera: Space Bunny Alpha Is Probably MiniMax M3.1-Flash