Anonymous Stealth Model "Space Bunny Alpha" Debuts on OpenRouter With 1M-Token Context, Free During Preview

OpenRouter has begun routing requests to Space Bunny Alpha, a new stealth model from an unnamed third-party provider. The model is available free of charge during its preview window and offers a 1M-token context window, according to its OpenRouter listing.

What's known

Space Bunny Alpha (stealth/space-bunny-alpha) supports native multimodal input and includes an adjustable reasoning-effort parameter, letting users trade off latency for deeper chain-of-thought processing. OpenRouter lists a full 1,000,000-token context window. The company behind the model has chosen not to disclose its identity during this preview phase, and OpenRouter states it is acting only as a routing layer — it is not the model's developer, owner, or operator.

Pricing is currently $0 per million tokens for both input and output, though this is explicitly a temporary preview arrangement rather than a permanent rate. Pricing not yet disclosed for post-preview access.

OpenRouter's own telemetry shows a P50 latency of 0.93 seconds and throughput of 79 tokens per second, best-case figures across providers. Reliability metrics are mixed: the model showed 100% uptime when requests were successfully routed to a provider over the trailing three days, but overall availability — which counts errors and empty responses — sat at 76.22% over that window, with OpenRouter-wide availability at 67.08% over the last 24 hours. These numbers suggest the backend serving Space Bunny Alpha is still unstable, consistent with an early preview deployment.

According to OpenRouter, prompts and completions sent to Space Bunny Alpha may be retained by the provider but are not used for training. All other usage is governed by OpenRouter's Stealth Model Terms.

Context: a pattern of stealth releases

Space Bunny Alpha is the latest in a series of anonymous model drops routed through OpenRouter under the "stealth" label. Two other current examples — Union Alpha and Ox Alpha — follow the same playbook: free preview access, anonymous branding, and eventual "reveal" once the free period ends. Union Alpha, for instance, was recently unmasked as Pareto, a production model from Unbiased, after its free stealth window closed. Ox Alpha, still active, is a 1M-context reasoning model aimed at coding and long-horizon agentic work.

This pattern lets model providers collect real-world usage data and community feedback without attaching a corporate name to early benchmarks or failure modes, while OpenRouter benefits from exclusive early access to unreleased models on its platform.

What this means

Space Bunny Alpha's claimed 1M-token context and multimodal support put it in the same tier as top-end frontier models on paper, but none of these capabilities have been independently benchmarked, and the provider's identity is unverified. The sub-77% availability figure is a bigger practical signal than any spec sheet claim: this is infrastructure still being hardened, not a production-ready endpoint. Treat performance claims as provisional until the model is deanonymized — as Union Alpha was — and tested against standard benchmark suites like MMLU, HumanEval, or SWE-bench under a known provider name.