On 9 June 2026, Anthropic released Claude Fable 5, the most capable model it has put in front of the general public so far.
Model launches happen often enough now that it's easy to tune them out. This one is worth a few minutes, because the thing that improved most is the exact thing that has kept most small businesses from using AI for anything beyond writing emails.
The headline: it can stay on a long job without losing the plot
Most AI tools are fine at short tasks. Ask for an email, a summary, a quick script, and you get something usable. Hand over a job with twenty steps and a dependency chain and older models tend to drift halfway through and quietly start making things up.
Fable 5's main improvement is stamina. It holds together across long, multi-step work, and the gap over older models gets wider the longer and messier the task is.
The example everyone's quoting is the Stripe one, so let me handle it carefully, because it sounds unbelievable until you read what it actually says. Anthropic reports that during early testing, Stripe used Fable 5 to run a codebase-wide migration inside a 50-million-line Ruby codebase, and that it finished in a day what would have taken a team over two months by hand.
Here's what that does and doesn't mean, because the headline version misleads people:
- It did not write or rewrite 50 million lines. That number is the size of the codebase it worked across, not the volume of new code. A codebase-wide migration is usually one well-defined, mechanical change applied in a lot of places, like renaming an API, bumping a framework version, or swapping out a deprecated pattern. Tedious, not clever.
- It was not unsupervised. Engineers scoped the change, set it up, and reviewed what came back before any of it shipped. The model did the bulk of the repetitive labour fast. It did not decide on its own what to do or merge its own work.
- Stripe is the best case, not the typical one. The reason a job like that compresses to a day is that Stripe has clean code, strong test coverage, and existing migration tooling. That maturity is the opposite of where most small businesses are.
So the honest read is: a tedious job that would have been two months of human grind got done in a day, with a person steering and a machine doing the donkey work. That's still a big deal. It's just not magic, and anyone who sells it to you as magic is the wrong person to be taking advice from.
You probably don't have a 50-million-line codebase. The point that survives is the principle. Work that used to need someone babysitting every step can now be handed over in bigger chunks, as long as someone defines the job and checks the result. For a lean business, that's often the difference between "we can't justify building that" and "let's try it this week."
We covered a similar pattern when Opus 4.8 shipped last month. Fable 5 moves the same needle further in the same direction.
Where it helps with the day-to-day
The gains aren't only for engineers. A few of the early results map onto things small businesses actually do:
Finance and analysis. On Hebbia's finance benchmark for senior-level reasoning, Fable 5 scored higher than any other model, with real gains in reading documents and interpreting charts and tables. If you've ever wanted a second set of eyes on a messy management report, that's the kind of task it now handles well. You still check its work, the same way you'd check a junior analyst's.
Contracts and review. One early tester, a legal team, ran a blind comparison and found Fable 5's redlines matched or beat their existing tool every time. For a small business without a lawyer on retainer, that's a useful first pass before you pay for the expensive second one. First pass, not final word.
Spreadsheets. Another tester found it beat the previous Claude model on everyday spreadsheet work at every setting and finished 25 to 30 percent faster. Reconciliations, data cleanups, the boring monthly stuff.
Reading what's on the screen. Its vision improved enough to rebuild a working web app from screenshots alone. The version of that you'll care about is pulling data out of invoices, forms, and scanned documents without building a fragile template first.
The price actually came down
This part surprised me. Fable 5 costs $10 per million input tokens and $50 per million output tokens, which Anthropic says is less than half the price of Mythos Preview, its previous top model. A more capable model that costs less is not the usual pattern. It means the cost of automating a given task keeps falling while the quality climbs, which is a good trend when you're the one paying the bill.
A few things to know before you lean on it
I'd be doing you a disservice if I only sold the upside. Three practical caveats.
The guardrails sometimes misfire. Because the model is capable enough to be dangerous in areas like cybersecurity and biology, Anthropic added filters that hand certain requests off to its older Opus 4.8 model instead. They've tuned these on the cautious side, so some harmless requests get caught too, though Anthropic says that happens in under 5 percent of sessions. If a security-adjacent task suddenly feels less sharp, that's probably why, and you'll be told when it happens.
Data handling changed. For this class of model, Anthropic now keeps 30 days of traffic, on its own platform and third-party ones. It says the data won't be used to train models and exists to catch attacks and reduce false alarms. If you're thinking about pushing sensitive client data through it, that's worth knowing first.
Access is being staged. Fable 5 is included on paid Claude plans at no extra cost through 22 June. From 23 June, using it needs usage credits until Anthropic has the capacity to fold it back into standard plans. So if you trial it now, check what your access looks like after that date before you build a workflow around it.
What this means in practice
Strip away the benchmarks and the story is simple. The work AI can finish with a person steering rather than babysitting just got bigger, and the price of that work went down. Capability you used to need a whole team or a serious budget to touch keeps getting cheaper and more reliable every few months.
The catch is the same as it's always been. A powerful tool doesn't help if nobody on your team has time to work out where it fits. The businesses that get value out of this aren't the ones with the best model. They're the ones who picked one or two real problems and wired the tool into how they already work. We covered what that looks like in our post on AI ROI in year one.
If you've got a process that eats hours every week, a pile of documents nobody wants to sort, or a "we should automate that" you've been putting off, we're happy to take a look and tell you honestly whether it's worth doing yet. Sometimes the answer is not now. When it is, we'll help you build it properly.
