Anthropic cuts AI model costs by up to 45 per cent in latest release

Anthropic has released its latest artificial intelligence models, Claude Fable 5.1 and Mythos 5.1, positioning the pair as direct responses to customer complaints regarding pricing, data retention, and overly strict safety measures. The company asserts that the new Fable 5.1 model delivers stronger performance than its predecessor while reducing costs by approximately 25 per cent for typical workloads. For complex agentic tasks that rely heavily on context reuse, the savings could reach up to 45 per cent, driven primarily by a significant reduction in the price of cached data.

The financial adjustment is particularly notable for enterprise users, with the cost of cache reads dropping by 75 per cent to $0.25 per million tokens. This pricing strategy aims to make the models more viable for high-volume operations where previously processed data is frequently accessed. While the underlying architecture for both Fable 5.1 and Mythos 5.1 remains the same, the two models differ in their availability and the level of safeguards applied to their outputs.

Early feedback from industry leaders suggests the update has improved practical utility. Dan Shipper, CEO of Every, described Fable 5.1 as the strongest coding model his team has used, noting its speed and token efficiency. Aaron Levie, CEO of Box, reported that his company’s agent, when paired with Fable 5.1, identified subtleties and ambiguities in data that the previous version had missed. Additionally, Lisan al Gaib observed that Mythos 5.1, when set to low reasoning, scores the same on benchmarks as its predecessor set to maximum reasoning, indicating a more efficient processing capability.

The release also addresses longstanding criticisms of Anthropic’s safety filters, which had previously generated false positives for benign content. Fable 5.1 features more precise safeguards that block 60 per cent fewer false positives, particularly for basic biology questions. In contrast, Mythos 5.1 retains the same biology restrictions as the previous model. Anthropic has also approved Fable 5.1 for identifying software vulnerabilities, although specific cybersecurity tasks such as penetration testing and exploit generation are still redirected to Opus models.

For enterprise clients concerned about data privacy, Anthropic has outlined its progress on data retention practices. The company’s Enterprise Frontier Safeguards, which store data on the customer’s own cloud servers rather than Anthropic’s infrastructure, are scheduled to roll out later this autumn. This feature is designed to offer complete privacy for sensitive corporate data.

In terms of market access, Claude Fable 5.1 is now available on all platforms, broadening its reach to the general public and enterprise sectors. However, Mythos 5.1 remains restricted to participants in the Project Glasswing programme and other trusted access initiatives, maintaining a tiered distribution strategy for the company’s most advanced offerings.