OpenRouter's Stealth programme is one of the most interesting — and frustrating — phenomena in the open AI ecosystem. Models with codenames like "Space Bunny Alpha" and "Ox Alpha" appear for free, for a few weeks, and then disappear. Nobody says who built them. Then, often, the truth comes out.
How the Stealth programme works
OpenRouter defines Stealth models as "developed and managed by a third-party provider that has chosen to remain anonymous during the preview". The provider uses OpenRouter's routing infrastructure; OpenRouter is neither the developer nor the owner of the model.
The typical characteristics:
- Free access for a limited period (often 1-2 weeks)
- Limited tokens but generous enough for testing on real traffic
- Hidden identity until the official "reveal"
- Specific terms of use: the Stealth Model Terms is a separate document from OpenRouter's standard terms, and it specifies that prompts and completions may be retained by the provider, but not used for training
The revealed models: who was really behind them
Of the fourteen Stealth models that appeared since April 2025, seven have been officially revealed — all through first-party channels (the OpenRouter blog or the lab itself).
Quasar Alpha and Optimus Alpha → GPT-4.1 and GPT-4 (OpenAI)
The first two stealth models, appearing in April 2025. Quasar Alpha was a pre-release snapshot of GPT-4.1, Optimus Alpha of GPT-4. OpenRouter revealed them two weeks later on its own blog.
Horizon Alpha and Horizon Beta → GPT-5 (OpenAI)
Appeared in July/August 2025 as "early GPT-5 checkpoints". Confirmed by OpenRouter itself in the post "GPT-5 is now live".
Ox Alpha → GLM-5.3-Flash (Zhipu AI / Z.ai)
August 2026. A model with a 1,048,576-token context, completely free. On 26 August, Z.ai officially revealed that Ox Alpha was their GLM-5.3-Flash, publishing the weights under an MIT licence the same day.
Union Alpha → Pareto (Unbiased)
A multimodal model for research and coding. The stealth period ended and it was replaced by Pareto, released by Unbiased.
Hunter Alpha and Healer Alpha → MiMo-V2-Pro (Xiaomi)
March 2026. Xiaomi confirmed that these models were internal builds of its MiMo team, led by former DeepSeek researchers.
Space Bunny Alpha: the mystery still unsolved
Space Bunny Alpha appeared on 23 September 2026 and is the most recent stealth model. It is free, has a 1-million-token context, supports multimodal input (text and images), advanced coding and adjustable reasoning effort.
Unlike the others, the provider has not yet revealed its identity. The community analysed the architecture and identified two main suspects:
MiniMax — The most credible hypothesis. Many engineers noticed that Space Bunny Alpha uses the same tokenizer as the MiniMax M3 family. MiniMax is an Asian tech company valued at billions of dollars, famous for ultra-fast, multimodal models with long contexts.
OpenAI — Some believe that "Space Bunny" may be an experimental version linked to the new GPT-6 family, because of the space-related name and the native handling of reasoning effort.
Neither hypothesis has been officially confirmed.
Stealth models in Hermes Agent
Hermes Agent, Nous Research's open-source autonomous agent, is the most common way to use Stealth models. On OpenRouter, Hermes Agent is the number one application worldwide for tokens processed, with over 53 trillion tokens in total.
Stealth models are accessible in Hermes Agent in two ways:
Via the interactive configurator:
hermes model
Select OpenRouter as the provider, enter the API key, and choose the desired model.
Via config.yaml:
model:
provider: openrouter
default: stealth/space-bunny-alpha
Hermes Agent's catalogue officially includes stealth/space-bunny-alpha and stealth/union-alpha as supported stealth models. stealth/ox-alpha is also routable, even if it may not appear in the picker (you can use an alias via /model <alias>).
Top models used in Hermes Agent on OpenRouter
According to OpenRouter's statistics, the most used models in Hermes Agent this month include Space Bunny Alpha with over 5.68 trillion tokens, positioned in second place behind a non-stealth model.
Why companies use Stealth models
The Stealth programme is not a simple experiment: it is a testing and marketing strategy.
- Validation on real traffic: companies can test their models on real users, with real workloads, before the official launch. This is invaluable for identifying bugs and failure modes.
- Zero cost for the user: being free attracts developers and power users who would otherwise never try a new model.
- Buzz and community: the mystery around the identity generates interest, discussions and technical analysis — exactly as happened with Space Bunny Alpha.
- Feedback: the provider can gather feedback from the community before launch, improving the final model.
The terms of use and privacy
The Stealth Model Terms are explicit: the provider "owns" everything users submit (prompts and completions), but commits not to use it for training. However, there is an important difference between "not used for training" and "deleted": the data can be retained, even if not used to improve the model.
Moreover, the terms specify that user content is licensed irrevocably to OpenRouter and its providers "to enable the provider to train, evaluate and improve models". There is a hashed identifier to prevent individual re-identification, but the legal structure is different from the standard terms.
For those with sensitive data, the rule is simple: do not send anything to a model whose provider you cannot identify or with whom you can enter into a DPA contract.
The future of the Stealth programme
The programme continues to grow. Since April 2025, fourteen stealth models have appeared, and the pace has increased. OpenRouter has made the programme increasingly structured, with clear terms, dedicated pages and a more defined reveal process.
For developers, stealth models represent a unique opportunity: accessing frontier models for free, for a limited time, with the chance to test them on real use cases before they become paid or disappear.
For those who wonder who built them — the answer, more often than you might think, comes within a few days. The reveal is almost always an official announcement, not a coincidence.
Sources and references: