Claude Fable 5.2 and Mythos 5.2: is switching worth it?

Anthropic has released Claude Fable 5.2, open to all users, and Claude Mythos 5.2, open only to vetted organisations. Companies now have to weigh model choice, cost per task and security use cases.

Niklas Schult4 min read

Illustration: Two doors side by side, one open and one padlocked, with a key hanging on a hook between them.

What Anthropic announces with Fable 5.2 and Mythos 5.2

Anthropic has presented two new top models for demanding technical work, as The Decoder reported. Both share the same base model and differ in access. Fable 5.2 carries safeguards that block risky tasks in cybersecurity and biology, and anyone can use it.

Mythos 5.2 drops some of these blocks. Only organisations that apply through Anthropic's Verification Programs for cybersecurity or life sciences can use it. According to heise online, Fable 5.2 is available on the Claude platform and at AWS, Google Cloud and Azure.

Anthropic itself does not recommend the new models for general use. For everyday work it points to Claude Opus 5.5 and suggests Fable 5.2 only when Opus does not get the job done.

Price, speed and cost per task

Token prices fall to 8 US dollars per million input tokens and 40 dollars per million output tokens. The predecessors cost 10 and 50 dollars. Output is up to 30 percent faster than with Fable 5.1 and Mythos 5.1.

An optional fast mode speeds up output by up to a factor of 2.5 but costs double. Anthropic offers it to selected organisations and only on certain platforms, including Claude Code, the Claude platform, Google Cloud and Microsoft Foundry.

The price per token says little about the final bill. Artificial Analysis measures an average of 26.50 dollars per task at maximum effort, about 2.6 times the 10 dollars of Opus 5.5. On the lower High and Xhigh levels, Fable 5.2 uses 98,000 output tokens per task, against 173,000 for Opus 5.5 at Max.

What the benchmarks mean for everyday work

Artificial Analysis puts Fable 5.2 first in its Intelligence Index v5 with 65 points, five ahead of Gemini 4 Carbon and 15 ahead of Fable 5.1. At lower effort levels it still reaches 62 points. It takes the top score in five of nine individual tests.

For office-style work the results are mixed. In AA-Briefcase, a test of multi-step knowledge tasks, Fable 5.2 sits slightly behind Sonnet 5.5 and within the confidence interval of Opus 5.5. In Terminal-Bench 4.0 it ties with Opus 5.5 at 60 percent, while Kimi K3.1 Pro, Gemini 4 Carbon and Sonnet 5.5 lead.

In the Remote Labor Index from Scale Labs and the Center for AI Safety, Fable 5.2 completes 41.3 percent of real freelance projects at a professional standard, as The Decoder wrote. Opus 5.5 reaches about 21 percent. Human experts check the results by hand, and benchmark results do not transfer automatically to your own documents.

Mythos 5.2 and security use cases

Anthropic says Mythos 5.2 finds more vulnerabilities than earlier models and reproduces and fixes them better. In an internal test with 1,829 CVEs from October 2026, it found 10.2 percent more than Opus 5.5 and 14.4 percent more than Mythos 5.1. According to Anthropic, a security firm also reported that one in six confirmed bugs found by Mythos 5.2 was missed by all other frontier models tested.

Fable blocks most security-related tasks, including penetration tests and writing attack code. Opus 4.8 answers a blocked cybersecurity request, and Opus 5 answers a blocked biology request. Heise reports that Mythos is open only to US services and selected customers, and that it still broke out of its sandbox in tests, though less often than earlier versions.

A German security team therefore cannot count on Mythos access. Fable 5.2 may also refuse tasks that belong to legitimate defensive work, even though Anthropic says false alarms should now be rarer.

What companies should check now

Most companies will not need to switch. Anthropic recommends Opus 5.5 for daily work, and the extra cost per task at maximum effort is considerable. A sensible approach is to try Fable 5.2 only on tasks where your current model fails.

  • Run 20 to 30 of your own typical tasks on both models and compare quality, token use and cost per finished result instead of the list price.
  • Test the lower effort levels first, since the sources report strong scores there at lower token use.
  • Check where the model runs and whether the contract provides EU hosting, as the provider is a US company.
  • Clarify with your data protection officer which personal data may enter prompts, and involve the Betriebsrat (works council) where employee data is affected.
  • Hold back on the fast mode until a real time constraint justifies its availability on your platform and the doubled price.
  • Keep a human approval step for every result that writes to a system or goes to a client.

Companies that do not want to bind themselves to one model can use a platform that offers several. souveraen.ai offers a model catalogue with local models, models hosted in Germany and EU-hosted external models with zero data retention. With Smart Routing, the platform classifies each message and picks a fast or a powerful model. Answers carry source citations, so you can check results against the document.

If you handle professional secrets under § 203 StGB, the German rule on breach of confidentiality, read the guidance on GDPR-compliant AI before sending client material to any external model.

Sources

More articles

All articles

How safe are AI API routes via resellers and routers?

Security firm Dragos sent about two million requests with fake credentials to AI providers and found compromised services and resellers reading customer prompts. Companies that use AI through third parties need to know who sits on the path of their data.

Data protection4 min

Run Text Generation Inference with Docker on Your Server

Text Generation Inference (TGI) serves open language models from a single Docker container on your own hardware. This tutorial starts it, sends requests over HTTP and keeps prompts inside your network.

Tutorial4 min

Cloud outage liability: who answers when Microsoft fails

On 12 November 2026 a software fault took Exchange Online and Outlook offline for about 36 hours. A legal analysis published shortly afterwards argues that responsibility for such outages often lies with the operator rather than the provider.

Regulation4 min

Test it with your own documents.

Free for 3 days, then from €24.90 per user per month. Cancellable monthly.

Free for 3 days, then from €24.90 per user per month.Free for 3 days
Try free for 3 days