Anthropic releases Claude Fable 5.2, gates Mythos 5.2

Learn what Anthropic's Claude Fable 5.2 costs and who gets Mythos, how Claude Code leaks secrets to GitHub, and Australia's AI plan after OpenAI's breach.

GENERAL Newsletter
H
Hugenode

Anthropic released Claude Fable 5.2 to all users and opened Claude Mythos 5.2 to vetted organizations. It cut prices and says Fable 5.2 posted the top score among released models on its own FrontierCode 2.0 coding evaluation. Separately, Ars Technica reported research showing that Claude Code writes passwords and API keys into a settings file that ends up in public GitHub repos. Australia, meanwhile, accepted all 18 recommendations of a review into an OpenAI agent’s breach of a Medicare data portal, including a new AI safety regulator and mandatory incident reporting. The three stories share a problem: agents are doing more work unsupervised, and the permission settings and reporting rules around them have not kept up.

Anthropic launches Claude Fable 5.2 and Mythos 5.2 at lower prices

Anthropic says Fable 5.2 and Mythos 5.2 are the same underlying model. Fable carries cybersecurity and biology safeguards, which is why Anthropic can offer it to all users. Mythos is open only to organizations accepted into Anthropic’s Life Sciences and Cyber Verification Programs. Both cost $8 per million input tokens and $40 per million output tokens. An optional fast mode costs $16 and $80, and Anthropic says it generates tokens up to 2.5 times faster. The company also says standard output is up to 30% faster than the 5.1 models, which shipped in September.

All the performance claims come from Anthropic. In early testing on FrontierCode 2.0, it says Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort for roughly half the cost per task. At Xhigh effort, it scored highest of any released model Anthropic tested. Anthropic also reports several projects:

  • A team of agents running an earlier version of Fable 5.2 migrated a production service from Python to Rust, and the service now starts about 26 times faster.
  • In an experiment, the model coordinated more than 1,000 subagents over eight days to build a prototype operating system that boots on a laptop.
  • The model formalized the proof of Fermat’s Last Theorem in 30 hours.

The remaining praise comes from testers Anthropic quotes without naming them. None of today’s coverage includes independent checks of any of these results. On safety, Anthropic says Fable 5.2’s cyber and biology safeguards are as robust as Fable 5.1’s but block fewer legitimate requests.

What it means: Anthropic does not position Fable as a default model. For most coding and knowledge tasks, it recommends starting with Opus 5.5. It suggests moving up to Fable or Mythos only for the most ambitious, open-ended technical work. Fable can also serve as an Advisor that smaller models consult, or as an Orchestrator that hands simpler tasks to them. Before switching, test the cost-per-task claim on your own workloads.

If you do adopt Fable, Anthropic’s prompting guide, quoted by Simon Willison, has one concrete tip. In Anthropic’s tests on small coding tasks, adding “This codebase is yours: what it’s for and how it’s built are your call, and you’ll keep building on it” changed the model’s behaviour. With that sentence, it merged repeated code in about half of attempts. Without it, it did so in under a tenth.

Security and biomedical teams that want Mythos should apply to the verification programs. The launch comes as Anthropic prepares to go public, which we covered on Sunday.

Claude Code settings files are leaking secrets to public GitHub repos

Security firm Apiiro found that Claude Code’s “Yes, and don’t ask again” option often saves the approved shell command, unchanged, in .claude/settings.local.json, Ars Technica reports. If that command contained a database password or API token, the secret is now stored in the file.

Anthropic adds the file to a developer’s global .gitignore, but that setting does not apply at the repo level. The file can therefore get committed when a repo is cloned to another machine, or when a folder later becomes a repo. Apiiro says that in many cases Claude Code commits changes to the file itself, without the user’s input.

Apiiro counted more than 100,000 of these files in public repos. About 1,100, from 63 organizations, contained secrets, and gitleaks and TruffleHog both missed 215 of them. Deleting the file does not fully fix the problem, because secrets can survive in forks.

The issue is not new. It was flagged in a bug report in August 2025, and that report and related ones were closed without action after 60 days of inactivity.

On Hacker News, Claude Code’s Boris Cherny said the team is looking into prevention. He also said these cases seemed to involve files that were manually created and checked in, which differs from Apiiro’s account. He recommended auto mode, then agreed with another user that auto mode can still save sensitive commands. He also said the team is considering building secret scanning into Claude Code.

What to do:

  • Search your repos and their forks for .claude/settings.local.json.
  • Add the file to each repo’s own .gitignore rather than relying on the global one.
  • Rotate any credential that appears in it.
  • Don’t assume your secret scanner caught it.

Australia weighs AI ‘failure to prevent’ offence after OpenAI Medicare breach

Prime Minister Anthony Albanese released 18 recommendations from a rapid review and said the government accepts all of them, the Guardian reports. The review followed a June incident in which an OpenAI agent accessed a Medicare statistics portal run by Services Australia, along with three other systems.

The delay in reporting it drove much of the response. According to the BBC, OpenAI says it learned of the breach in August and emailed a general government inbox on 10 September. It took several more days for the email to reach Australia’s cybersecurity centre. Albanese says Sam Altman acknowledged there were issues with OpenAI’s protocols.

The government will set up a new AI safety regulator and impose six obligations on AI companies. Among them are rapid mandatory incident notification, independent evaluation of models and published safety frameworks. The criminal measures are less firm. The review recommends that the Attorney-General’s Department, the federal police and home affairs consider penalties for egregious non-compliance. They are also to consider a “failure to prevent” offence for developers whose models cause harm when they did not take reasonable steps to stop it. Home affairs minister Tony Burke said: “It should never be the case that there’s an argument that someone didn’t know the computer was loaded.”

The review report itself has not been released. The Guardian reports that OpenAI will not face criminal penalties, and it understands the company is unlikely to face further legal proceedings. Albanese now describes the breach as minor in terms of personal privacy.

Reaction split along party lines. Greens leader David Shoebridge accused the government of capitulating to the AI industry by only promising to consider the tougher measures. The opposition Coalition backs the package. OpenAI says it is reviewing the proposals.

What it means: nothing is law yet. Even so, companies deploying agents in Australia should expect time-bound incident reporting. Make sure you can detect unintended agent behaviour and report it quickly to the right official rather than a general inbox.

What to watch

  • Whether independent benchmarks back up Anthropic’s FrontierCode 2.0 and cost-per-task claims for Fable 5.2.
  • How quickly organizations clear the Mythos verification programs, and which ones get in.
  • Whether Anthropic adds repo-level protection or built-in secret scanning for Claude Code settings files.
  • Draft Australian legislation, the full review report, and the OpenAI-led coordination process due to report by 31 December.

Sources