Anthropic's strongest publicly available model, built on Mythos-class intelligence with safeguards active. Safeguards fire in less than 5% of sessions; affected queries fall back to Opus 4.8. Mythos 5 is the same model with some safeguards lifted, deployed through Project Glasswing for authorized cyberdefenders and infrastructure providers.

SWE-Bench Pro: 80.3% (next model: 69.4%). FrontierCode Diamond: 29.3% vs 13.4% for Opus 4.8. Spatial reasoning: 38.6% vs 14.5%. Pricing: $10/$50 per MTok input/output. Free on Pro, Max, Team, and Enterprise through June 22, then usage credits.

Adds Fable 5 access. Fixes session transcript saving for sessions launched from VS Code integrated terminal or shells with inherited Claude Code environment variables (they now appear in --resume).

Safe mode, directory change, and bundled skills controls. The new --safe-mode flag (and CLAUDE_CODE_SAFE_MODE env var) disables all customizations (CLAUDE.md, plugins, skills, hooks, MCP servers) for clean troubleshooting. /cd moves a session to a new working directory without breaking the prompt cache. disableBundledSkills setting hides bundled skills and slash commands from the model.

Bug fixes: enterprise MCP policy enforcement on reconnect, Up/Down arrow navigation in long wrapped lines, macOS UI stall for claude.ai-authenticated users, Windows slash-command scan slowness, Remote Control stuck on reconnecting after session resume, Git Credential Manager popups on Windows startup, stale permission prompts on remote reattachment, background agents ignoring project-level env values, MCPB plugin cache spurious invalidation on Windows, untrusted project settings bypassing OTEL trust confirmation. Reduced CPU during streaming. Restored 5-minute idle timeout on Vertex/Foundry.

A Swift package that makes Claude available as a server-side language model in Apple's Foundation Models framework. Conforms to the LanguageModel protocol so developers can drive it with LanguageModelSession alongside Apple's on-device model. Supports multi-step reasoning, code generation, web search, code execution, and SwiftUI streaming. Targets iOS 27, iPadOS 27, macOS 27, visionOS 27, watchOS 27. Announced June 9, available June 10.

Bug fix: adds frontier_llm refusal category.

Adds support for Managed Agents deployments and environment variable credentials.

Adds claude-fable-5 and claude-mythos-5 model identifiers. Server-side fallbacks on refusal. Client-side fallbacks middleware for providers that do not support server-side fallbacks.

Bug fix: adds frontier_llm refusal category.

Adds support for Managed Agents deployments and environment variable credentials.

Adds claude-fable-5 and claude-mythos-5 model identifiers. Server-side fallbacks on refusal. Client-side fallbacks middleware. ctx.logger in middleware. Fixes 3p middleware ordering.

All four platform TypeScript SDKs fix 3p middleware ordering, released simultaneously at 16:36 to 16:37 UTC.

GitHub Copilot subscribers can select Claude Fable 5 as their model in the Copilot interface as of June 9, 2026.

Fable 5 is available through Amazon Bedrock. Reported by AWS as "Mythos-class capabilities with built-in safeguards."