Ability to disable models would be a great feature for multiple reasons. Personally, I do not trust OpenAI or Grok - specially when they are planning to have ads in there.
Is this in the pipeline or something that can easily implemented?
Hey! Thanks for this great question.
You can do this today with the Agent API. It accepts either a preset or an explicit models array, and that array doubles as an ordered fallback chain — so if you pass models=[, "perplexity/sonar", "anthropic/claude-sonnet-4-6"], the runtime will only ever try those two, in that order.
One thing to know: presets do route to specific third‑party models by default (for example, pro-search defaults to GPT‑5.x under the hood). To enforce your own policy but still keep a preset’s tools and system prompt, you can override the model alongside the preset, e.g.:
client.responses.create(preset="pro-search", model="anthropic/claude-sonnet-4-6", input=...). The presets docs explicitly show that all preset parameters, including model, are overrideable.
If you’d rather stay Sonar‑centric, the Agent API also has Sonar‑style options: you can just use a Sonar model ID in models (or as model) and build your fallback chain around that, or stick to a single Sonar model only. The Agent API is designed as the “multi‑provider” layer for the platform, while the Sonar API is the more focused, web‑grounded chat endpoint.
If you’re migrating from the Sonar API, the shape changes a bit (messages → input, and search filters/tools move under tools[...]), but the mental model is similar and the migration is pretty straightforward. The Sonar quickstart and Agent API quickstart have concrete examples for both sides if you want to compare them line–by–line.
If you have any further questions, we’d love to answer them and keep the discussion going.