Verified against the June 9, 2026 announcement
Claude Fable5 specs & model card
Every number that matters, on one page — context window, output limits, the effort parameter, safeguard behavior, and pricing. Dated and sourced; updated whenever Anthropic changes anything.
The spec sheet
All figures from Anthropic's official launch documentation, June 9, 2026.
| Spec | Value |
|---|---|
| API model ID | claude-fable-5 |
| Release date | June 9, 2026 |
| Tier | Mythos-class — priced a tier above Opus |
| Context window | 1M tokens |
| Max output | 128K tokens |
| Thinking | Adaptive, always on — no separate extended-thinking toggle |
| Effort parameter | low · medium · high (default) · xhigh/max |
| Sampling controls | temperature and top_p removed |
| API pricing | $10 / $50 — per million tokens (input / output) |
| Prompt caching | 90% off cache hits; minimum cacheable prefix 2,048 tokens |
| Batch API | 50% off input and output; stacks with caching |
| Safeguard fallback | Under 5% of sessions answered by Opus 4.8, billed at Opus rates |
| Data retention | 30 days for safety monitoring — zero-retention terms do not apply |
| Availability | Claude API, Claude apps, AWS, Google Cloud, Microsoft Foundry, GitHub Copilot, OpenRouter |
Source: Anthropic announcement and model documentation, June 9, 2026. Spotted a change? Tell us via the support widget.
The effort parameter
Effort replaced temperature as the main quality / latency / cost control. Four levels:
01
low
Fastest and cheapest. Routine transformations, classification, light chat — work where Haiku-grade output would already pass.
02
medium
Balanced. Day-to-day coding assistance and drafting where latency matters more than the last few points of quality.
03
high
The default. Full reasoning depth for serious engineering and analysis work — most production agent configs run here.
04
xhigh / max
Maximum depth. Frontier-hard debugging, final-pass reviews, capability-sensitive work. Slowest and most expensive — reserve it for tasks that earn it.
One model, two releases
Fable5 shares weights with the restricted Claude Mythos 5 — what you can buy is the safeguarded variant.
Claude Fable5
The public release. Strict safety classifiers; cybersecurity and biology requests may be quietly answered by Opus 4.8 (under 5% of sessions), billed at Opus rates.
Claude Mythos 5
Same weights, safeguards lifted in some areas. Restricted to vetted cyberdefenders and infrastructure providers through Project Glasswing — not purchasable.
Specs FAQ
Quick answers on Claude Fable5's technical specifications.
Build on Claude Fable5 with your eyes open
Know the Claude Fable5 cost math, the safeguard behavior, and the right tier before you commit. Create a free account to get notified when specs, prices, or access channels change.