Claude Fable 5 Is Here — and It's Built to Work for Days Without You

On June 9, 2026, Anthropic released Claude Fable 5 — the public version of the model that had been circulating internally under the name 'Mythos.' It's the company's fifth-generation model and its most powerful one available to the public. The headline isn't a benchmark number. It's the shift in what the model is built for: long-running, asynchronous work. Anthropic says Fable 5 can operate continuously for days when run inside an agent framework, sustaining complex tasks that previous models couldn't hold together.

That's the real change. Earlier models were things you prompted — you asked, they answered, you steered. Fable 5 is built to be delegated to. You give it a multi-stage job, and it works through it: writing code, running its own tests, validating its output against the original goal, and pushing forward with minimal oversight. It's less 'chatbot' and more 'autonomous worker you check in on.'

Three areas stand out. Software engineering: large codebase migrations, complex implementations, and multi-day autonomous coding sessions where it writes its own tests and checks the result against the spec. Knowledge work: multi-stage research-to-deliverable projects it can run with little hand-holding. And vision: reading diagrams, charts, and tables inside PDFs, and evaluating coding output against the original design.

The numbers back it up. Anthropic reports Fable 5 scoring more than 10% higher than Claude Opus 4.8 on some benchmarks. The analytics platform Hex said it was the first model to hit 90% on its core analytics benchmark. Base44 reported it was superior at 'one-shotting' full apps with strong tool-calling. Early testers also noted 25–30% faster execution on spreadsheet tasks and stronger results on finance and legal analysis. Anthropic says at least 95% of Fable sessions run entirely on the model itself without falling back to anything else.

This is where Fable 5 gets unusual. The model has hard limits in high-risk domains — cybersecurity, biology, chemistry, and distillation. If a request gets flagged in one of those areas, Fable blocks its own response and quietly reroutes the query to the older, more constrained Claude Opus 4.8. You don't get charged Fable's premium price for those rerouted requests. Anthropic stress-tested the system through external bug bounties — over 1,000 hours with no universal jailbreaks found — plus red-teaming before release.

There's also a privacy trade-off worth knowing: Anthropic imposed a mandatory 30-day data retention window on all Fable traffic, even for enterprise customers who previously had zero-retention agreements. The company says the retained data is used only to defend against novel attacks and catch false positives — not for training. The context matters too: Fable 5 launched just days after Anthropic publicly warned that frontier AI systems may soon be capable of recursive self-improvement, and urged the major labs to coordinate safeguards on how fast they push. Shipping their most powerful public model while warning about the danger of powerful models is the tension Anthropic is openly living in right now.

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — roughly double Claude Opus 4.8. The 90% input-token discount for prompt caching still applies, and US-only inference carries a 1.1x multiplier. Anthropic is temporarily including Fable 5 in existing subscriptions through June 22, after which access moves to consumption-based pricing.

On availability, it's broad from day one: natively on the Claude Platform, on the consumption-based Enterprise plan, and through Amazon Bedrock, Google Cloud, and Microsoft Foundry. So whatever stack you're already on, you can probably reach it without changing providers.

Here's the part that matters for builders. A model that can work for days unattended changes what one person can ship. The bottleneck stops being 'how fast can I prompt and steer' and becomes 'how well can I define the job and let it run.' For solo builders and small teams, that's the same leverage shift I've been talking about for months — except the model now sustains the long, boring, multi-step work that used to need a human babysitting every step.

But the doubled price and the safety routing mean you don't just default to Fable for everything. The smart move is to match the model to the job: Fable 5 for the hard, long, high-value work where its reliability pays for itself, and cheaper, faster models for routine calls. And if you build in security, health, or anything regulated, understand the guardrails before you architect around it — the automatic fallback behavior changes what you can and can't ship. Powerful is good. Knowing when to reach for the powerful one is the actual skill.