Claude Fable 5: Anthropic's most powerful AI yet, and what it means for your business

Every few months a new "best AI model" arrives and the headlines blur together. This one is worth a closer look, because the gap is unusually large.

On an independent test built by the team at Every that measures whether a model can do the work of a senior software engineer, Claude Fable 5 scored 91 out of 100. Claude’s own previous flagship, Opus 4.8, scored 63. GPT-5.5 scored 62. When the gap between last month’s best and this month’s best is nearly 30 points, something real has changed.

What Anthropic actually launched

Fable 5 is the first public release of Anthropic’s top tier, internally called "Mythos-class". The pitch is not a faster chatbot. It is a model built to run long, complex tasks on its own, the kind that used to fall apart halfway through because the AI lost the thread.

• It can work autonomously for days on a single task, staying focused across millions of words of context.
• It writes and tests its own code to check it actually works.
• It is strong at vision, reading charts and scientific figures, and even rebuilding a web app’s code from a screenshot.
• It costs $10 per million words of input and $50 per million of output, roughly double the previous flagship.
• For risky topics (cyber-attacks, bioweapons, chemistry), it quietly hands the question back to the older, safer Opus 4.8. That happens in fewer than 5% of sessions.

How good is it, really?

Benchmarks are easy to game and vendors pick the flattering ones, so treat the numbers as a direction, not gospel. With that caveat, the direction is striking:

• Coding: top of Cognition’s FrontierCode evaluation, even at "medium effort", and the best model on Every’s vibe-coding test.
• Analytics: the first model to break 90% on Anthropic’s core analytics benchmark, a ten-point jump over Opus.
• Finance: the highest score of any model on Hebbia’s finance benchmark.
• Research: on a physics problem it reached in 36 hours roughly where GPT-5.5 landed after four days, using a third of the computing effort.

The most telling number is not a benchmark at all. Stripe, the payments company, says it used the model to run a 50-million-line code migration in a single day, work it reckons would have taken a team more than two months by hand. That is the shape of the change: not "a bit better at answers", but "weeks of skilled work compressed into a day".

The model race, in one line

Here is the thing to hold onto while the leaderboards churn: the model is becoming a commodity that keeps leaping ahead. Fable beats Opus, which beat the one before it; in a few months something will beat Fable. The price keeps falling and the capability keeps climbing. Tying your business to one model, or panicking each time a new one lands, is the wrong reflex. We made that case in full when we wrote about the difference between a tool and a managed service.

What it makes possible for a smaller business

You are not going to run a 50-million-line migration. So what does a model this capable actually change for a company of 5 to 200 people?

The unlock is work that runs unattended. Until now, useful AI mostly meant a person prompting a chatbot and pasting the answer somewhere. A model that can hold a long task together on its own makes a different thing possible: an agent that reconciles last month’s accounts, drafts the board pack from your own numbers, chases the slipping deal, or works through a backlog overnight, and has a person sign it off in the morning. That is the shift from AI as a tool you operate to AI as a result that lands on your desk, which is exactly the direction we wrote about in when AI builds itself.

The practical catch: this capability is most valuable when it is pointed at your data and your processes, safely. A frontier model that cannot see how your business runs is just a very clever stranger. One that is wired into your systems, under your control, is leverage. That is the whole idea behind a private, custom AI model.

The safety footnote that is really the headline

There is an uncomfortable detail worth noticing. Anthropic released Fable 5 only days after warning publicly that AI is becoming dangerously capable. The most powerful version, "Mythos 5", with the safety guardrails removed, is not on sale; it is restricted to government cyber-defenders under a programme called Project Glasswing. The public Fable model ships with classifiers that block misuse and fall back to the older model when a question strays into dangerous territory.

For a business owner the lesson is not fear, it is posture. Capability is now outrunning the controls around it. The sensible response is the boring one: keep a human accountable for anything that matters, keep sensitive data under your control and under UK rules, and do not pour confidential information into whatever public model is topping the charts this week. (If that is on your list, our free UK AI policy template covers it.)

What to do about it on Monday

Nothing dramatic. Specifically:

1. Do not switch everything to the new model. Use the strongest model only where the task is genuinely hard, and a cheaper one for the rest.
2. Pick one workflow where "weeks of work in a day" would actually change your month, and test it there.
3. Keep a person in the loop before anything acts on its own.
4. Stay model-agnostic, build so you can swap in whatever tops the list next quarter.

Claude Fable 5 is a real jump, not just a new label. But the winners will not be the businesses that chase every release. They will be the ones that quietly point a capable model at the work that drains their week, and keep it under control. That is what we build at Nerdster: private, capable AI wired into the tools you already use, run by a UK team, with a person always accountable.