Anthropic launched the Fable 5.1 and Mythos 5.1 AI models on Tuesday. The headline is not a new benchmark. It is a price line that changes how teams build. Cache reads now cost $0.25 per million tokens, down from $1.00. That is a 75 percent cut.
Fable 5.1 is more capable yet cheaper than Fable 5, thanks to lower pricing on cached data. Input and output pricing stays at $10 and $50 per million tokens. Cache writes also stay put, at $12.50 for five minutes and $20 for one hour. The savings show up when an agent reuses the same context again and again. Anthropic estimates the lower cache price cuts typical workload costs about 25 percent and highly agentic workloads by as much as 45 percent.
That math matters for anyone running long sessions. Think codebases that stay in memory. Think research threads that span hours. Think customer support bots that keep a user's history warm. Every cache hit is now a quarter of what it was. The bill bends.
Mythos 5.1 is designed for cybersecurity and life sciences work and will initially be available only to members of Anthropic's trusted access programs. It is the same model under the hood as Fable 5.1, but with safeguards tuned for those domains. Anthropic said it updated the cybersecurity and biology safeguards so they intervene less often on benign requests. Cyber safeguards now intervene about 60 percent less per session. Biology safeguards intervene 85 percent less often on harmless questions.
Wider rollouts of the new models are planned for this fall. For now, Mythos 5.1 is limited to vetted US organizations through two programs, the Cyber Verification Program and the Life Sciences Verification Program. Anthropic plans to expand access to international partners over time. Fable 5.1 is generally available today on cloud platforms and through the Anthropic API.
Anthropic is refining safeguards, expanding cloud-based privacy options, and allowing selective security vulnerability detection. The company introduced a new security architecture called Enterprise Frontier Safeguards, or EFS, designed to let organizations retain monitoring data inside infrastructure they control. That addresses a common complaint about data retention policies. Teams want the power of a frontier model without sending every log to a third party. EFS is a step in that direction.
I like this move. It shows the company listened. Cost, data retention, and safety constraints were the big three complaints about the previous models. The response is not a press release. It is a price cut and a control panel.
There is a catch. The base input and output prices did not move. $10 and $50 per million tokens are still frontier rates. The savings only materialize if your workload hits the cache hard. If you are sending fresh prompts every time, you will not feel the 75 percent cut. You will feel the same bill.
Another catch. Mythos 5.1 is not open to everyone. It is a gated model for a narrow set of users. That makes sense for safety. It also means a small team working on a novel bio or security project may not get access right away. They will have to apply, wait, and prove they are vetted. The timeline for international expansion is vague. Over time could mean months.
Fable 5.1 and Mythos 5.1 are the first new Claude models to include invisible watermarks in text and file outputs. That is a quiet change with big implications. It helps trace model outputs back to their source. It also raises questions about how those watermarks behave in downstream systems. Will they survive copy-paste? Will they break when converted to PDF? Anthropic did not share those details on Tuesday.
The pricing inversion is interesting. Cache reads on Fable 5.1 now cost 2.5 percent of the base input price, rather than the 10 percent multiplier used by most other Claude models. That makes agentic Fable 5.1 workloads cheaper per session than Opus 5 in some cases. It is a signal that Anthropic is optimizing for a specific pattern of use. Long-running agents that keep context warm. Not one-off queries.
This release addresses industry demands for cost-effective and secure AI solutions. It could influence market competition and adoption in critical sectors like cybersecurity and life sciences. Cheaper cache reads lower the barrier for teams that want to experiment with persistent agents. Enhanced security features make it easier for regulated industries to try frontier models without violating internal policies.
I am optimistic about what this enables. Smaller teams can now afford to keep more context in memory. They can build agents that remember more and cost less to run. That opens up new product ideas. It also makes advanced AI more accessible and safer for a wider range of applications.
But I have been burned before by pricing that looks great on a slide and falls apart in production. The 25 percent and 45 percent savings are estimates. They depend on your cache hit rate. They depend on your token mix. They depend on how well your architecture reuses context. If you are not careful, you will pay for cache writes at $12.50 or $20 per million tokens and not get enough hits to make it back.
The wider fall rollout is also a variable. Anthropic has a history of rolling features out in stages. What is available today on the API may not be available on every cloud platform until later. Teams planning migrations need to check their provider's timeline.
I am watching two things next. First, how fast Mythos 5.1 access expands beyond the initial US-only programs. Second, how the EFS controls work in practice. If they let enterprises keep data on their own infrastructure without killing performance, that is a real unlock. If they add latency or complexity, teams will stick with the default.
The new models promise to make advanced AI more accessible and safer for a wider range of applications. That is the optimism I feel reading this announcement. The catch is that the savings and the safety are conditional. They depend on your workload, your provider, and your access status. Wonder without the catch is advertising. The catch is what makes the wonder trustworthy.
Explore: Startup