Anthropic just shipped Claude Fable 5, and a single number tells you why it matters: $10 per million input tokens for a model that, according to early customers, does in 36 hours what GPT-5.5 does in four days.
Agents That Work for Days, Not Hours
Fable 5 is built for asynchronous, long-running tasks. Drop it into an agent harness like Claude Code or Claude Managed Agents, and it plans across stages, delegates to sub-agents, and checks its own work—running for days without human hand-holding. Anthropic explicitly calls out large migrations, complex implementations, and multi-day autonomous sessions as target use cases.
That self-checking loop is the key. The model writes its own tests, uses vision to compare outputs against design goals, and kills incorrect beliefs mid-stream. One customer (a major legal firm) reported that Fable 5's redlines matched or beat their current model in blind review every time.
Benchmarks That Show Real Gains
Fable 5 is the first model to break 90% on Anthropic's core analytics benchmark for complex, long-running analytical tasks—a 10-point jump over Opus 4.8. On ViBench (end-to-end vibe-coding), it nearly saturates base use cases while using fewer tokens. On CursorBench, it's state-of-the-art.
Perhaps the most telling result: in frontier physics research, Fable 5 used a third of the reasoning tokens to get nearly to GPT-5.5's output after four days—in just 36 hours. On everyday spreadsheet work, it beats Opus 4.8 at every effort level while finishing runs 25–30% faster.
Pricing, Availability, and the Safety Fallback
Fable 5 costs $10/M input tokens and $50/M output tokens, with the existing 90% discount on prompt caching. It's available on the consumption-based Enterprise plan, through Anthropic's own platform, AWS, Google Cloud, and Microsoft Foundry. US-only inference is available at 1.1x pricing.
Anthropic built in a safety fallback: queries flagged for cybersecurity or biology are automatically routed to Opus 4.8, and customers aren't charged Fable prices for those rerouted requests. Data retention is 30 days for safety monitoring.
Fable 5 doesn't just move the needle on token efficiency—it changes what kind of problems you can even attempt to solve autonomously. The extra thinking pays for itself when you can hand it a week's worth of work and check the results once.
Source: Claude Mythos 5 / Fable 5
Domain: anthropic.com
Comments load interactively on the live page.