What changed, the one quirk worth knowing, and whether it's worth switching from your daily driver.
Fable 5 is designed to hold a complex, multi-step job together over hours — and to pick work back up across sessions without you re-explaining everything. If you've ever watched a long task fall apart at step six, this is the model Anthropic built for that.
Anthropic trains Fable 5 to answer when it's confident and decline when it isn't, rather than inventing facts or sources. They even score it that way: a confident wrong answer counts against the model, and a well-placed "I'm not sure" counts better than a guess. But in their own internal testing, the model still sometimes stated an unverified guess as fact — and when a user pushed back, it went and checked, found the test behind its claim had never run, and said so plainly. The lesson for your work isn't "trust it blindly" — it's the opposite: a confident-sounding answer can still be wrong, and one "are you sure?" is the cheapest quality check you have.
— the failure and the recovery are both documented in Anthropic's Fable 5 system card
Fable 5 ships with extra built-in safeguards around high-risk technical topics like cybersecurity and biology. In practice: if your question lands in one of those areas, Claude automatically handles it with Opus 4.8 instead.
What this means for you: if you work in security or life sciences and notice some answers feel different from others, that's the safeguard working — not a bug. For everyone else, you'll likely never notice.
Stay on Sonnet 4.6 if your Claude work is writing, summarizing, meeting prep, analysis, and everyday documents. It's still the daily-driver default for a reason.
Reach for Fable 5 when the job is genuinely big: a multi-hour research pass, a project that spans many documents and steps, work where being told "I'm not sure" beats a confident guess.
Under the hood, Fable 5 handles up to 1 million tokens of context — hundreds of thousands of words — and can produce very long outputs (up to 128,000 tokens) in a single pass.
This generation of the model is Anthropic's most resistant yet to "prompt injection" — hidden instructions buried in a webpage or document that try to hijack an AI assistant. That's a real improvement. But the habits don't change: don't paste secrets, be careful with links from unknown sources, only connect tools you trust. Better locks, same habits.