Add structured output support for Anthropic provider

[vojto]

Hey, I tried to support structured output for Anthropic.

The only problem is that RubyLLM::Schema currently adds strict: true to each schema it creates. Here's a PR in that repo to fix that: danielfriis/ruby_llm-schema#28

Until that is merged, the workaround is to create this subclass of RubyLLM::Schema:

class AnthropicSchema < RubyLLM::Schema def to_json_schema result = super result[:schema].delete(:strict) result[:schema].delete("strict") result end end 

---

AI generated:

Summary

• Implements structured outputs using Anthropic's structured-outputs-2025-11-13 beta API
• Adds structured output capability detection for Claude 4+ models
• Includes comprehensive test coverage for the new functionality

Changes

• Anthropic Provider: Added complete method override to inject the structured-outputs beta header when schema is provided
• Capabilities: Added supports_structured_output? method to detect Claude 4+ models that support structured outputs
• Chat: Implemented output_format parameter with json_schema type in payload rendering
• Tests: Added specs for beta header handling and output_format payload generation
• Refactoring: Extracted claude3_or_newer? and claude4_or_newer? helper methods for cleaner version detection

Test plan

• Added unit tests for beta header handling
• Added unit tests for output_format payload generation
• Verified structured output capability detection for Claude 4+ models
• Manual testing with actual Anthropic API (recommended)

🤖 Generated with Claude Code

[tpaulshippy]

Do you plan to add or modify specs that hit the real API and produce new/updated VCR cassettes?

[vojto]

@tpaulshippy that should be done now.

I edited models.json manually, because I don't have all the API keys to run rake models:update. I'm assuming someone else will update it after merging the PR?

I modified models_to_test.rb a little - made a separate array for models that we wanna test for schema. I added haiku-4.5 and sonnet-4.5 to this list, and kept gpt-4.1-nano and gemini-2.5-flash that was already there.