Skip to the cards
Filed
Source
Thomas Smith
Loans
986

Claude Fable 5.2 and Mythos 5.2 Cut API Prices 20% as Coding and Vulnerability Scores Rise

Anthropic’s latest models cost less and report stronger coding results, but supervision, data retention and Mythos access controls remain key deployment questions.

Claude Fable 5.2 and Mythos 5.2 Cut API Prices 20% as Coding and Vulnerability Scores Rise

Anthropic’s Claude Fable 5.2 and Mythos 5.2 pair lower API rates with reported gains in coding and vulnerability discovery. For IT teams, that may improve the economics of long-running software tasks, but the headline price per token won’t settle the deployment decision. Reliability, data handling and oversight still shape the cost and risk of putting autonomous agents to work.

Both models share an underlying model, while their safeguards and availability differ. Fable is broadly available. Direct access to Mythos is limited to vetted cybersecurity and life sciences organizations, a restriction Anthropic says is intended to limit potential misuse. Claude Enterprise customers can also access Mythos vulnerability detection through Claude Security without joining those verification programs.

Lower rates change the cost calculation

Anthropic has set both models at $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the previous versions. The company also says the new models can generate output up to 30% faster than their predecessors. Together, lower rates and faster generation could reduce the expense and waiting time associated with some extended tasks.

Task completion matters as much as token pricing. Lian Jye Su, chief analyst at Omdia, said improved performance could make long-running AI agents more viable for software migrations and scientific research. If a model finishes more assignments successfully, organizations may spend less on repeated attempts or human rework. That benefit will vary with the task, the amount of oversight required and the model’s effort setting.

Anthropic says Fable 5.2 at its “High effort” setting achieved a coding benchmark score comparable to Opus 5.5 at its highest setting, at roughly half the cost per task. That comparison may make Fable a candidate for demanding work where teams are weighing model quality against spend. It doesn’t establish a universal saving: actual costs depend on workload and configuration.

For most coding and knowledge work, Anthropic recommends Opus 5.5. Fable and Mythos are positioned for more demanding assignments. Sanchit Vir Gogia, chief analyst at Greyhound Research, summed up the budgeting issue this way: “Anthropic has cut the price of a token, not the price of a thought.” He urged CIOs to track AI processing effort and compare spending with successfully completed tasks. Teams evaluating the new rates will need to measure end-to-end results, rather than treating a lower token price as a full measure of savings.

Benchmark gains need operational checks

Fable 5.2 scored 49.1% on the FrontierCode 2.0 benchmark at its Xhigh effort setting, compared with 32.4% for Fable 5.1. The benchmark measures whether maintainers of open-source projects would accept AI-generated code changes. It gives teams a specific point of comparison for coding performance, though the score alone doesn’t show how the model will behave in a company’s own repositories, review process or legacy applications.

An internal Anthropic experiment offers a more unusual example of sustained work: an 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 could boot on a laptop. This shows the scale of task the company tested, but it doesn’t establish that an enterprise can delegate a comparable production project without supervision.

Su said the performance gains bring organizations closer to highly autonomous software development, while reliability and coordination remain obstacles. “Reliability remains an issue without human supervision,” he said. Managing agents across company tools and teams is also difficult. In practice, businesses considering these systems will need visibility into agent actions, monitoring, and appropriately isolated environments. Critical changes should receive human review, and records of agent behavior should support later audits.

Pareekh Jain, CEO of Pareekh Consulting, pointed to complex legacy systems and the need to meet business requirements as limits on how much development work companies can delegate. For administrators, that means assessing integration and review requirements alongside benchmark performance. A model that handles a discrete coding task may still need substantial direction when a change touches older systems or multiple internal teams.

Mythos vulnerability results raise a dual-use question

Anthropic reported that Mythos 5.2 identified 89.6% of vulnerabilities in its internal CVE Rediscovery evaluation, up from 78.3% for Mythos 5.1. The evaluation used 1,829 previously disclosed vulnerabilities. It tested whether models could independently rediscover flaws in software snapshots taken before those issues were fixed. Anthropic also said Mythos 5.2 improved at reproducing vulnerabilities and developing fixes.

Su said, “Mythos 5.2 shifts the balance towards defenders in controlled enterprise settings.” Jain cautioned that better automated discovery could also help attackers find software weaknesses faster, describing an “AI-powered race between attackers and defenders.” The reported score is therefore relevant to defensive teams assessing vulnerability detection, but access restrictions and controlled use remain central to how Anthropic is offering the model.

Retention terms belong in procurement review

Anthropic says customers eligible for zero data retention can continue using both models under those arrangements. Other customers face 30-day retention for safety monitoring. The company has also introduced Enterprise Frontier Safeguards, currently in private beta, which lets automated safety monitoring run on prompts and responses stored within a customer’s infrastructure.

Those terms can affect whether a workload is suitable for the service. Gogia said data retention requirements could outweigh the benefit of lower API rates for regulated enterprises, and urged procurement teams to obtain written confirmation of zero-retention arrangements before using the models for sensitive workloads. IT teams should assess eligibility, retention terms and monitoring arrangements alongside performance tests and cost estimates. The practical question is whether the models complete the organization’s target tasks reliably enough to justify their usage and oversight costs.

Source: Thomas Smith · www.infoworld.com

Filed alongside