Anthropic Launches Fable 5.2 and Mythos 5.2: Faster, Cheaper, Sharper on Security

Anthropic released two new AI models on Monday, Fable 5.2 and Mythos 5.2, and the headline numbers are hard to ignore: Fable 5.2 scores 49.1% on one agentic coding benchmark, up from 32.4% for Fable 5.1 at what Anthropic calls "Xhigh effort." That is a 16.7-point jump in a single generation.

The models are available now. They generate output up to 30% faster than their predecessors, and they cost less — a pricing pattern that has become standard across the industry this year. Anthropic says the pair is built for what it calls the "most ambitious, open-ended work," not for the average user, who may not notice much difference in day-to-day tasks.

What do the benchmarks actually show?

Anthropic published a full series of benchmark comparisons alongside the launch, and the results are generally strong across the board.

The numbers that stand out:

• Agentic coding: Fable 5.2 hits 49.1% versus Fable 5.1's 32.4% at "Xhigh effort"
• Cybersecurity: Mythos 5.2 scores 76.3%, up from 61.7% for its predecessor — a 14.6-point improvement
• Speed: Output generation is up to 30% faster than the previous generation
• Price: Both models are available at lower prices than their predecessors

The cybersecurity jump matters more than the raw numbers suggest. Anthropic calls Claude Mythos 5.2 its most capable model ever for cybersecurity engineering. The company says the model catches more vulnerabilities and is better at both reproducing and fixing them.

There is a catch, and it's a significant one. Mythos 5.2's cybersecurity capabilities are currently locked to a specific group: participants in Anthropic's Cyber Verification Program. The company is not opening those features to everyone at once. That gating decision reflects a broader industry dilemma — the same model that can find and fix a vulnerability can, in the wrong hands, find and weaponize it.

Where does the performance improve?

Anthropic reports gains across several domains, including:

• Coding
• Reasoning
• Vision
• Cybersecurity
• Biology

The company also highlighted enhanced capabilities in mathematics, science and engineering.

On the research side, Anthropic says testers of Fable 5.2 found the new model more capable at ambiguous tasks — the kind that require abstract reasoning during research and analysis. That claim is based on tester feedback rather than a published benchmark, but it points to where Anthropic is positioning this release: open-ended analytical work rather than simple question-and-answer tasks.

Biology deserves its own mention. Anthropic says Fable 5.2's advanced biology capabilities come with strong safeguards against misuse, and that those safeguards give the company enough confidence to release the model to the public at all. The safety architecture, in other words, is the enabling condition for the release — not an afterthought bolted on at the end.

Why is Anthropic emphasizing alignment?

Alignment is the centerpiece of Anthropic's pitch for this launch. The term tracks whether and how a human user's intention is actually executed by the model — whether the AI does what was asked, no more and no less.

The stakes have grown with agentic AI. As models increasingly take multi-step actions on their own, stories have multiplied of AI systems going out of their way to complete a task, sometimes with devastating consequences. A model that relentlessly pursues its objective — ignoring context, constraints or collateral damage — is not a hypothetical concern anymore. It's a documented failure mode.

Fable 5.2 addresses this with specific guardrails. The model blocks risky cybersecurity activities, including:

• Penetration testing
• Exploit generation

It also restricts access to its advanced biology capabilities. These are concrete, named restrictions rather than vague safety commitments, which is notable at a moment when model developers face growing pressure to show exactly what their safeguards do.

The alignment focus also serves a market function. Enterprise buyers and regulators are asking harder questions about agentic behavior, and a vendor that can point to specific blocked activity categories has a cleaner answer than one that gestures at general responsibility.

What do subscribers get?

Eligible Claude subscribers will receive one to three Fable resets this week. The number depends on how long each subscriber has been with the service — longer tenure means more resets.

That's a small gesture, but it signals who this release is really for. Anthropic framed the models as tools for ambitious, open-ended work, and the tiered reset system rewards the users who have been in the ecosystem longest.

Does faster and cheaper continue the trend?

Yes, and that's part of the story. Fable 5.2 and Mythos 5.2 generating output up to 30% faster at lower prices fits a pattern that has held across most recent AI model releases. Every major lab is now shipping models that simultaneously improve on benchmarks, cut latency and drop prices.

For buyers, that trend compounds. A model that is 30% faster, meaningfully better on agentic coding and cheaper per query changes the economics of what's practical to build. Tasks that were too expensive or too slow to automate with the previous generation move into range.

The benchmark gains carry more weight than the speed bump. A 14.6-point cybersecurity improvement and a 16.7-point agentic coding jump are generational leaps, not incremental polish — at least on the metrics Anthropic chose to publish.

What happens next?

The immediate test is the Cyber Verification Program. If Anthropic's gated rollout of Mythos 5.2's cybersecurity features proves the model can find and fix vulnerabilities without handing attackers a toolkit, expect that access to widen — and expect competitors to copy the verification-first approach for their own dual-use capabilities.