Space Bunny Alpha: 1M Tokens in Stealth Mode

What OpenRouter has actually launched

What stands out here is not the anonymity, but the combination of specifications: Space Bunny Alpha is listed with a one-million-token context window, fast inference, strong coding capabilities and native multimodal input. OpenRouter’s stealth-model page says it accepts text, images and video, while returning text.

The model is available through OpenRouter’s unified API under the identifier stealth/space-bunny-alpha. During the stealth preview, prompts and completions were priced at zero. This is not a separate protocol or an exotic wrapper: integration looks like selecting a standard model in a familiar chat-completions workflow.

Adjustable reasoning effort makes it possible to trade speed for answer depth. That is useful for code and agent workflows, where applying the same reasoning mode to every step only burns time. However, no formal benchmark table for Space Bunny Alpha appears in the available materials, so claims about speed and coding remain OpenRouter specifications rather than independently measured results.

There is also a less appealing part of the specification. OpenRouter explicitly states that the model belongs to an anonymous third-party provider and that prompts and responses may be retained, although they are not used for training. For proprietary code, internal documents and personal data, this is not fine print; it is the first architectural constraint.

A million tokens alone guarantees nothing. I would first test fact retrieval across the full context, degradation in long conversations, latency on large inputs and the consistency of structured output. A large window matters only when the model truly preserves the relevant relationships instead of merely accepting an enormous payload.

What changes for developers

The main change is straightforward: costly experiments with long and multimodal context became more accessible during the stealth preview. A single request can potentially contain a large codebase, documentation or a mixed set of text and visual materials. That reduces the need for upfront data chunking, but it does not remove the need to validate retrieval quality.

Free access is also convenient for prototypes and comparison runs. Still, it should not underpin a product’s long-term economics: the model’s developer is undisclosed, while pricing and terms after the stealth period are not defined in the provided materials. Research speed improves, but operational predictability declines.

My engineering conclusion remains cautious: this is not empty hype, because the combination of one million tokens, multimodality and adjustable reasoning is genuinely interesting. But Space Bunny Alpha’s value will be determined not by window size, but by long-task quality, retention policy and behavior after the provider is revealed. For now, the model’s most important parameter is hidden not in the API, but in its origin.

We previously covered Pony Alpha on OpenRouter, including its 200K context window and its use for low-risk model testing. That comparison helps frame what Space Bunny Alpha adds to OpenRouter’s growing model catalog.