OpenRouter quietly launched two stealth models named Sonoma (Dusk and Sky versions), touting "maximum intelligence" and a 2M-token context window—8 times longer than GPT-4 Turbo's official limit.


Free comes at the highest cost. While completely free during testing, the fine print clearly states: all prompts and outputs will be recorded for model improvement. This mirrors xAI's Grok playbook—users confirmed it's Grok in disguise through system tags.


Notable details:

1. Models claim to be created by "Oak AI," but leftover xAI tags in the codebase contradict this

2. Exceptionally fast responses (4.28s latency), matching Grok 1.5's characteristics

3. Supports image input and parallel tool calls, but programming capabilities reportedly lag behind Gemini

The model is already integrated with AnyCoder and a dozen other dev tools, with real-time usage visible in VS Code extensions. But free lunches have strings attached—users report hitting usage limits after 4-5 messages despite the "completely free" claim.

Such "stealth testing" is becoming common. Technically, blind testing via third-party platforms yields authentic user feedback; ethically, "free" services that hide data usage terms essentially make users pay with their data.

If trying it, enable "Allow first-party model data logging" in privacy settings. Better yet, avoid discussing trade secrets or personal data—when even the model's true identity requires code-sleuthing, data pipelines are even less transparent.