Glossary — Fable 5 terms

Updated July 1, 2026. Fable 5 is back after the June export-control suspension; new entries below cover the suspension-and-return vocabulary. Latest news →

FableGuide › Glossary

Fable 5: Anthropic's most intelligent model (ID claude-fable-5), released 2026-06-09 — a new tier above Opus. Suspended worldwide 2026-06-12 under a US export-control directive, restored globally 2026-07-01 with a new safety classifier. Source.
Mythos 5: The same underlying model with safeguards lifted in specific domains. Restricted to vetted partners — currently US organizations only (restored 2026-06-26); see Mythos, explained and the announcement.
Mythos-class: Anthropic's term for this capability tier — powerful enough that public versions ship with active classifier safeguards rather than training-time restrictions alone. Source.
Adaptive thinking: thinking: {"type": "adaptive"} — the model decides when and how deeply to reason per request. The only thinking mode on Fable; fixed budget_tokens is gone.
Effort: output_config.effort — how hard the model works: low, medium, high (default), xhigh, max. Scales thinking depth, tool-call consolidation, and verbosity together.
Thinking display: Whether reasoning text appears in responses. Defaults to omitted on Fable; set display: "summarized" to show users a summary while it thinks.
Task budget: A beta feature giving the model a visible token allowance for a whole agentic run, which it paces itself against. Distinct from max_tokens, the hard per-response cap it can't see. Recipe 06.
Structured outputs: Schema-guaranteed JSON via output_config.format or messages.parse(). The replacement for assistant prefills, which now return 400. Recipe 03.
Prompt caching: Reusing the unchanged front of your prompt at ~10% of input price. Prefix-match by exact bytes; Fable's minimum cacheable prefix is 2,048 tokens. Recipe 04.
Classifier safeguards: Runtime filters watching cyber, bio/chem, and distillation topics. Anthropic reports they trigger on under 5% of sessions; independent testing has suggested ~8–9%. Distinct from refusals trained into the model itself.
Fallback (Opus 4.8): When a safeguard trips, the response is handled by Claude Opus 4.8 instead of Fable. You still get an answer — from the safer sibling.
Distillation: Training a competing model on another model's outputs. Fable's third classifier watches for extraction attempts.
Project Glasswing: The partner program whose members get Mythos 5 with cyber safeguards lifted, for security research. Since the June 2026 restoration it is also the vehicle for expanding Mythos access beyond US organizations. Source.
Export control directive: The US government order of 2026-06-12 that forced Anthropic to suspend Fable 5 and Mythos 5 for all users worldwide — it required restricting foreign nationals, and with no real-time nationality verification Anthropic pulled access entirely. The Commerce Department lifted the controls 2026-06-30. Details.
Safety classifier (July 2026): The new runtime classifier Fable 5 returned with on 2026-07-01, targeting the reported cyber jailbreak; it blocks the technique in over 99% of cases. Its deliberately expanded margin produces false positives on routine coding, sometimes handing the task to Opus 4.8 with a notification. Details.
Jailbreak severity framework: Anthropic's proposed industry scoring for jailbreaks on four criteria: capability gain over existing tools, breadth of offensive tasks enabled, ease of weaponization, and discoverability. The most severe trigger immediate preliminary mitigations; Anthropic wants it codified in regulation and applied equally across frontier developers.
Subscription window: Originally: Fable included free on claude.ai paid plans 2026-06-09 → 06-22, cut short by the 06-12 suspension. Replaced by the return window: 50% of weekly usage limits 2026-07-01 → 07-07, usage credits after. Details.
Return window: The post-restoration make-good: Pro, Max, Team, and select Enterprise plans include Fable 5 at half of weekly usage limits from 2026-07-01 through 07-07, then usage credits. Details.
Agents First: Joshua Baer's framework (agentsfirst.dev): build for the agent that decides as well as the human who pays. It's why this site has /llms.txt and typed /api/ twins.

The Fable vocabulary