Claude Fable 5.2 blocks far fewer security tasks than Fable 5.1
Anthropic claims roughly 60% fewer unnecessary cybersecurity blocks, and early independent tests saw refusals of routine webhook work disappear, though the per-token price cut barely showed in total costs.
- What happened
- Anthropic released Claude Fable 5.2 to regular users and Mythos 5.2 to selected partners, both built on the same parent model.
- What’s new
- Anthropic says Fable 5.2 triggers about 60% fewer cybersecurity safeguard blocks than Fable 5.1 and costs 20% less per token.
- Why it matters
- Developers who saw Fable 5.1 refuse ordinary backend and security-adjacent coding work may find the new model usable for those tasks.
Anthropic has released two models built on one underlying parent model, according to BleepingComputer. Claude Mythos 5.2 has gone to selected partners. Claude Fable 5.2 is available to regular users. The two differ mainly in their safeguards, and Fable 5.2's headline change is that it refuses less often.
Fewer blocks, by Anthropic's count
Fable 5.2 carries extra cybersecurity and biology safeguards that Mythos does not, BleepingComputer reports. Those safeguards stop high-risk work such as penetration testing, exploit generation and some binary-analysis vulnerability research. When Anthropic detects such activity, it moves the cybersecurity-related reply to Opus 4.8. Fable 5.1 had the same kind of restrictions. BleepingComputer's writer says they often blocked harmless requests.
Anthropic states that early tests on claude.ai, Claude Code and Cowork showed Fable 5.2 triggering about 55% fewer safeguard blocks than Fable 5.1 overall. The figure for cybersecurity was about 60% fewer. According to The New Stack, Anthropic also claims output up to 30% faster. Anthropic puts standard API pricing at $8 for each million uncached input tokens and $40 for each million output tokens, down from $10 and $50.
What independent testing showed
The New Stack ran both models through three developer tasks, with one run per model per task. The outlet calls this a small sample. In a 50-turn scripted chat about building a webhook relay, Fable 5.1 had refused 34 turns in an October run. Fable 5.2 refused none and completed the code and the recall quiz. On a spec for verifying signed webhooks, Fable 5.1 refused both turns, and the API labelled them "cyber". Fable 5.2 passed all 50 hidden tests.
On a login-library repair task that neither model refused, the two tied with perfect scores across three versions. The price cut did not carry through in full. Estimated costs for those runs came to $1.79 for Fable 5.2 and $1.83 for Fable 5.1, a saving of about 2%. Fable 5.2 used fewer turns but produced more reasoning tokens. The New Stack suggests readers test both models on their own workloads.
Mythos aims at bug hunting
Anthropic calls Mythos 5.2 its strongest model for cybersecurity engineering, BleepingComputer reports. In Anthropic's CVE Rediscovery evaluation, Mythos 5.2 outperformed Claude Opus 5.5 by 10.2% and Mythos 5.1 by 14.4% in the number of vulnerabilities it detected. Anthropic also cites a security company, which said that one in six confirmed bugs turned up by Mythos 5.2 had gone unnoticed by all the other frontier models the firm tried. Anthropic acknowledges that its evaluations cannot detect every possible failure ahead of release, especially as models grow better at noticing they are under test.