Anthropic has lowered API rates by 20% for Fable 5.2 and Mythos 5.2, but the number on the price sheet will tell you little about your next AI bill. What you pay depends on how much effort you ask the model to spend, how often tasks finish without rework and whether your compliance rules block the cheaper path. Teams planning agent budgets for 2027 should model cost per completed task, not cost per token.
What happened
According to InfoWorld, Anthropic released Claude Fable 5.2 and Mythos 5.2. Both run on the same underlying model with different safeguards. Anthropic offers Fable broadly, while it restricts direct Mythos access to vetted cybersecurity and life sciences organizations. Both cost $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the previous versions. Anthropic says output is up to 30% faster.
Anthropic reported that 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 asks whether open-source maintainers would accept AI-generated changes. The company also said Fable 5.2 at High effort reached a coding benchmark score comparable to Opus 5.5 at its highest setting, at roughly half the cost per task. Anthropic still recommends Opus 5.5 for most coding and knowledge work.
On security, Anthropic said Mythos 5.2 found 89.6% of vulnerabilities in its internal CVE Rediscovery test of 1,829 disclosed flaws, up from 78.3% for Mythos 5.1. Claude Enterprise customers can use those detection capabilities through Claude Security without joining the verification programs. SiliconANGLE describes the two models as identical in core capability, with Fable blocking certain biology and cybersecurity use cases.
Why it matters
A price cut paired with a large benchmark jump could change the economics of long-running agents. Omdia’s Lian Jye Su told InfoWorld that the combination could make agents viable for software migrations and scientific research, and that higher completion rates could cut deployment costs further. In an internal experiment, an Anthropic engineer had Fable 5.2 coordinate more than 1,000 subagents over eight days to build a prototype operating system that could boot on a laptop.
The caveats are just as clear. Greyhound Research’s Sanchit Vir Gogia said Anthropic cut “the price of a token, not the price of a thought”. He urged CIOs to treat processing effort as a budget item, which suggests a cheaper rate can still produce a larger bill if you run at demanding settings. Su added that reliability without human supervision remains a problem and that coordinating many agents across enterprise tools is hard. Pareekh Jain of Pareekh Consulting pointed to legacy systems and business requirements as limits on how much work enterprises can delegate.
The security side cuts both ways. Su sees a shift toward defenders in controlled settings, while Jain warned that better vulnerability discovery also helps attackers.
What it means for you
Rebuild your cost model. Run your own representative tasks at several effort settings and record spend per accepted result, including human review time. Anthropic’s half-cost comparison with Opus 5.5 comes from its own benchmark testing, and the savings will vary by workload and effort setting. Our AI Tool Radar tracks live pricing across 30 AI tools if you need a baseline for comparison.
Pick the model by task difficulty. Since Anthropic itself points most coding and knowledge work to Opus 5.5, route routine jobs there and reserve Fable for the hard migrations and long-horizon work. Set effort caps so one runaway agent cannot burn through a monthly budget.
Check data retention before you celebrate. Customers eligible for zero data retention can keep it, but others face 30-day retention for safety monitoring. Gogia argues this can outweigh the price benefit for regulated firms and urges procurement to get written confirmation of zero retention before sensitive workloads go live. Ask about the private-beta Enterprise Frontier Safeguards, which run monitoring on prompts and responses stored in your own infrastructure.
Keep humans in the loop. Run agents in isolated environments, log their activity for audits and require approval for critical changes. If your agents grow past what serverless can handle, our piece on when an AWS agent outgrows Lambda offers a workload test for the next step.
What to watch
- Independent results on FrontierCode 2.0 and real-world task costs, which will test Anthropic’s own benchmark claims.
- Whether rivals answer with their own price cuts or effort-based pricing.
- Wider availability and terms for Enterprise Frontier Safeguards beyond private beta.
- Evidence on whether stronger vulnerability discovery helps defenders faster than attackers.
