Morning Edition · Sunday, November 22, 2026Published at 1:11 AM EST · New York
Its restricted counterpart, Claude Mythos 5.2, runs the same underlying model with fewer safeguards and scores 89.6 percent on a benchmark for rediscovering known vulnerabilities.

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9. It called them its most capable models for coding, finance and science. Nearly two weeks of third-party testing since then show where the gains are real and where the headline numbers need caveats.
The biggest reported jump is on FrontierCode 2.0. Fable 5.2 scores 49.1 percent at its highest effort setting, up from 32.4 percent for Fable 5.1. At Medium effort it scores 37.6 percent (VentureBeat). That spread shows much of the gain comes from spending more inference tokens. At High effort, Fable scores 45.3 percent, close to Opus 5.5's 45.5 percent at maximum effort (Lindy). Lindy also reports 69.9 percent on FrontierSWE v2, against 62.3 percent for Opus 5.5. Anthropic's robotics-simulation score rose from 27.6 percent to 64.9 percent (Yahoo Tech).
Some results need qualification. Fable's maximum Terminal-Bench score counts graded results from attempts that timed out. Excluding those attempts lowers one result from 45.7 percent to 44.4 percent (VentureBeat). On HealthBench, one analysis puts Fable 5.2 at 58.8, below Opus 5.5's 64.3 (Aleinikov). The new model does not improve on every task.
For security teams, Mythos 5.2 is the more important release. Anthropic restricts access to it. It shares Fable's underlying model but has fewer safeguards. It scored 89.6 percent on Common Vulnerabilities and Exposures (CVE) Rediscovery, up from 78.3 percent for Mythos 5.1 (Yahoo Tech). Because both products share the same weights, restricting access, rather than how the model was trained, now determines who can use its offensive-security capability.
What this means
Fable's coding gains depend heavily on effort settings, so token budget, not just model choice, now sets the cost of agentic coding. That shifts spending toward heavy-use coding customers. …
The desk expects OpenAI and Google to answer with code-specific releases within months. Independent FrontierCode results below Anthropic's numbers would weaken Anthropic's lead.
Part of a tracked trend
Frontier Labs Race on AI Coding Capability
Coding is becoming a primary competitive battleground among frontier labs, with incumbents standing up permanent coding teams and investing in new training stages (e.g. midtraining) to match leaders like Anthropic; expect recurring reorganizations, benchmarks, and model releases aimed specifically at code.
Start a discussion in Townsquare.
More from this edition
What to watch
Observations to monitor, not financial advice.
Synthesized from: Anthropic News · VentureBeat · Yahoo Tech · Lindy · Aleksei Aleinikov
Comments
0No comments yet.