Anthropic quietly enables Claude Cowork to run on any AI model—GPT-5, Grok, local models, or enterprise clouds—with no official announcement

Anthropic shipped third-party inference support for Claude Cowork and Code in Claude Desktop without a blog post or announcement, discovered by accident on April 22, 2026. Users can now point Claude Desktop at GPT-5, Grok, Gemini, open-source models via OpenRouter, local models, or enterprise gateways (Amazon Bedrock, Google Vertex AI, Azure Foundry). All admin controls (token caps, tool allowlists, auto-update blocks) work identically across individual and enterprise setups.

Instead of being locked to Anthropic's own AI model, Claude Cowork now works as a platform: the same agentic workspace (with Code tab, skills, plugins, and MCP servers—tools that connect to services like Jira or Confluence) runs behind any LLM. An individual on OpenRouter can test Cowork for free using tencent/hy3-preview; an enterprise on Bedrock keeps data inside its compliance boundary while using the same interface.

For individuals hitting Claude's $100–$200/month usage limits, this eliminates the subscription wall—you can now use Cowork with free or cheaper models. For enterprises blocked from sending data to Anthropic by security policy, this lets them deploy the full Claude Desktop harness (agents, automation, MCP integrations) without routing through Anthropic's servers. For AI model providers like OpenRouter or OpenAI, this turns Claude Desktop into a potential distribution channel for GPT-5 and other models.