Anthropic's top new AI does the same coding work for about half the cost
Anthropic released Claude Fable 5.2 and its security twin, Mythos 5.2, on 9 November. The new model takes on longer, harder jobs with less hand-holding, and costs less than the version before it.
The short version
- Anthropic released a new version of Claude Fable on 9 November. It is the most capable model the company sells to everyone.
- On a test of real coding work, it matched the best score of Anthropic's everyday model for roughly half the cost per task.
- A twin version, Claude Mythos, keeps stronger hacking skills and goes only to vetted security and life-science teams.
What Anthropic released
Fable and Mythos are the same AI model with different safety settings. Fable blocks most hacking and biology research requests, so Anthropic can sell it to anyone. Mythos keeps those skills and is open only to approved organizations.
The new version answers up to 30% faster than the one before it, and it costs less to use.
Anthropic says the model beats Opus, its strongest everyday model, on almost every test it ran. Even so, it tells most users to start with Opus. Fable is for the hardest and most open-ended work.
Fable also wrongly blocks fewer harmless requests than before. That matters if your staff found the old version refusing ordinary work.
What it did in testing
Anthropic's examples are long jobs that used to need a team of people.
- Working with over 1,000 copies of itself for 8 days, it built a working computer operating system from scratch.
- From a one-paragraph request, it designed three working electronic devices, circuit boards included. Each board was ready to manufacture in under a day.
- A team of its helpers rewrote a key company service in a faster programming language. The service now starts about 26 times faster.
Outside testers reported business wins too. Box, a document software company, said the model caught salary deposits in a loan application that did not match the bank statement. Oracle said the model got its database totals right, while two other models overcounted by 5 to 8 times.
What it means for Indonesia
The model is on all the major cloud platforms. Indonesian companies can buy it through the cloud account they already have.
The bigger change is price. On Anthropic's coding test, the same quality now costs about half as much. That makes long, many-step projects easier to justify, such as moving old software to new systems.
The way people use these models is changing too. Anthropic describes Fable handing simpler pieces of a job to smaller, cheaper models. It can also act as an adviser that smaller models consult on hard problems. For a company, that means one big job can mix expensive and cheap AI, and the bill depends on how well it is split.
Banks and insurers should look at the Box example. Checking that loan papers agree with each other is slow work for people. This is the kind of check the model handled well in Box's test.
Security teams should watch the Mythos side. Anthropic says a security company found that 1 in 6 confirmed bugs from Mythos were missed by every other leading model it tested. Similar skills may reach attackers through other tools.
What to do next
- Ask your teams which AI projects stalled because the work was too long or too costly. Price them again.
- Keep an everyday model as the default and send only the hardest jobs to Fable, as Anthropic itself advises.
- If you handle customer data, check where your prompts are kept. Anthropic keeps them for 30 days for safety checks unless you qualify for zero data retention.
- Ask whoever tests your systems how they plan for AI that finds software flaws this well.