Fable 5, Mythos 5, and the one agent that got upgraded
Every claim on this page was verified against Anthropic's live documentation and announcement on 2026-07-01, the day this site shipped -- sources linked at the bottom. Nothing here is recalled from model memory.
Claude Fable 5: what's verified
- What it is. Claude Fable 5, API model ID
claude-fable-5, is Anthropic's most capable widely-released model, described as "next-generation intelligence for long-running agents." - How it thinks. Adaptive thinking is always on -- there is no manual
budget_tokensparameter to set. - Window sizes. 1M-token context window; 128k-token max output.
- Availability. Generally available since 2026-06-09 on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
- The suspension. Both Fable 5 and its sibling Claude Mythos 5 were unexpectedly suspended for a period (roughly 2026-06-12 to 2026-06-30) while Anthropic worked with the U.S. government on an export-control review. Access was restored 2026-07-01, with Fable 5 becoming available that day to users globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork.
Claude Mythos 5: explicitly not used here
Claude Mythos 5 (API model ID claude-mythos-5) is not general availability. It is invitation-only, limited to approved customers inside Anthropic's "Project Glasswing" (for defensive cybersecurity and life-sciences work), with no self-serve signup. This repo does not use Mythos 5 anywhere -- AgentModel in crates/schema/src/agent.rs has no variant for it, no seed row in this site's database references it, and this build correctly does not attempt to. A model catalog that listed an ID nobody here can call would be exactly the kind of invented content this family exists to avoid.
What this repo did with those facts
On restoration day, one agent definition changed. engineering-coworker is the only one of the nine that actually executes operations -- real wrangler deploys, real git commit && git pushes, real D1 SQL -- on the developer's Mac, via the engineering-coworker MCP binary. Highest stakes, longest horizon; "long-running agents" is its literal job description. So it moved to Fable 5, in the typed source first (crates/schema/src/agent.rs gained the Fable5 variant, crates/agent-gen/src/agents.rs switched exactly one agent to it) and then regenerated via cargo run -p agent-gen.
| Before 2026-07-01 | After | |
|---|---|---|
| model | claude-haiku-4-5-20251001 | claude-fable-5 |
| tools | Bash Read | Bash Read -- unchanged |
| max_turns | 15 | 15 -- unchanged |
| description | unchanged -- the trigger text was already right; only the engine changed | |
Just as deliberately, the other eight agents did not move. The Fable5 variant's own doc comment in the schema is the standing policy: "Reserve for agents doing genuinely high-stakes, long-horizon work -- not a drop-in replacement for Haiku on mechanical tasks." cargo-check does not need next-generation intelligence to run cargo check and count warnings.
Sources
- platform.claude.com/docs/en/about-claude/models/overview -- model catalog: ID, context window, max output, adaptive thinking.
- platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5 -- the 2026-06-09 introduction, including Mythos 5's Project Glasswing scoping.
- anthropic.com/news/redeploying-fable-5 -- the 2026-07-01 restoration announcement.