Anthropic cuts Claude API prices 20% and ships Fable 5.2 and Mythos 5.2

Anthropic has cut API prices by 20% with the launch of Claude Fable 5.2 and Mythos 5.2, while claiming material gains on complex software engineering tasks. Input tokens now cost $8 per million and output tokens $40, down from $10 and $50 respectively. The company also says the new versions generate output up to 30% faster than their predecessors.

The two models share one underlying model but carry different safeguards. Fable 5.2 is broadly available; direct access to Mythos 5.2 is restricted to vetted cybersecurity and life sciences organizations, a deliberate brake on misuse.

What do the benchmarks actually show?

Fable 5.2 scored 49.1% on the FrontierCode 2.0 benchmark at its Xhigh effort setting, against 32.4% for Fable 5.1. The benchmark measures whether open-source project maintainers would accept AI-generated changes to their code — a usefully brutal standard.

Anthropic also ran a price-performance comparison. Fable 5.2 at its "High effort" setting matched a coding benchmark score of Opus 5.5 at its highest setting, at roughly half the cost per task. The savings depend on workload and effort setting, so treat the half-cost figure as a ceiling rather than a promise. Anthropic still recommends Opus 5.5 for most coding and knowledge work, reserving Fable and Mythos for demanding assignments.

What does this change for enterprise budgets?

Plenty, according to Lian Jye Su, chief analyst at Omdia. Improved performance and lower prices could make long-running AI agents viable for software migrations and scientific research, and higher task completion rates could cut deployment costs further.

Sanchit Vir Gogia, chief analyst at Greyhound Research, put a sharper point on it: "Anthropic has cut the price of a token, not the price of a thought." He urged CIOs to treat AI processing effort as a budget line and to measure costs against successfully completed tasks, not raw token consumption.

How autonomous can these agents get?

In an internal experiment, an Anthropic engineer asked Fable 5.2 to build a prototype operating system. The model coordinated more than 1,000 subagents over eight days and produced software that booted on a laptop. Eight days, a thousand subagents, one bootable result.

Su credits the improvement with moving enterprises closer to highly autonomous software development, with a caveat. "Reliability remains an issue without human supervision," he said. Coordinating multiple agents across enterprise tools and internal teams stays difficult. Organizations should monitor agent activity, run the systems in isolated environments, keep critical changes subject to human approval, and retain enough visibility into agent behavior to support audits.

Pareekh Jain, CEO of Pareekh Consulting, added a business-side constraint: complex legacy systems and business requirements will cap how much development enterprises can delegate to AI agents.

Can defenders outrun attackers?

Mythos 5.2 identified 89.6% of vulnerabilities in Anthropic's internal CVE Rediscovery evaluation, up from 78.3% for Mythos 5.1. The test covered 1,829 previously disclosed vulnerabilities and checked whether models could independently rediscover flaws in software snapshots taken before each fix. Anthropic says the model also improved at reproducing vulnerabilities and developing fixes.

Access matters here. Direct Mythos 5.2 use requires Anthropic's verification programs, but Claude Enterprise customers can use its vulnerability detection through Claude Security without joining them. Su's verdict: "Mythos 5.2 shifts the balance towards defenders in controlled enterprise settings."

Jain counters that faster AI vulnerability discovery cuts both ways. Attackers can use the same techniques to find weaknesses more quickly. "This creates an AI-powered race between attackers and defenders," he said.

What about data retention?

Anthropic will let customers eligible for zero data retention keep that arrangement on both models. Everyone else faces 30-day retention for safety monitoring. The company has also launched Enterprise Frontier Safeguards, now in private beta, which runs automated safety monitoring on prompts and responses stored inside the customer's own infrastructure.

For regulated enterprises, Gogia argues retention terms could outweigh any savings from cheaper APIs. His advice to procurement teams is blunt: obtain written confirmation of zero-retention arrangements before deploying these models on sensitive workloads.