Anthropic did not cut the sticker price on its newest model. It cut the thing that actually drives the bill. Claude Fable 5.1, released on September 1 alongside the restricted Claude Mythos 5.1, keeps base pricing at 10 dollars per million input tokens and 50 dollars per million output tokens. But the cost of cache reads, the charge for having the model pull on information it has already processed, drops 75 percent, from 1 dollar to 25 cents per million tokens. For typical workloads Anthropic estimates roughly a 25 percent saving. For long-running agent tasks, the company says the figure can reach 45 percent.
That distinction matters more than any benchmark chart. An AI agent working through a codebase or a stack of documents does not start fresh with every step. It keeps re-reading the same system instructions, tool definitions, files and conversation history, and each of those re-reads is a cache hit. Make cache hits nearly free and the economics of running an agent for hours instead of minutes change completely.
Two names, one model
Fable 5.1 and Mythos 5.1 are the same underlying system. The difference is who can use it and with which safeguards. Fable is generally available and, according to Anthropic, improves on the June release across coding and multi-step knowledge work while delivering similar or better results at lower reasoning settings. Mythos remains limited to vetted organisations working in fields like cybersecurity and life sciences, where the model’s capabilities are considered too sensitive for open access.
The split is becoming a template. Fable 5.1 can now help identify vulnerabilities in source code, but the more dangerous end of that skill, generating working exploits, stays behind the Mythos gate. OpenAI, for its part, said this week that its forthcoming Astra model has crossed the Critical cyber capability threshold under its own Preparedness Framework, triggering extra controls before any broad deployment. Two rival labs are arriving at the same conclusion: the frontier now has to be sold in tiers.
Fewer refusals
The other headline change is about friction. Anthropic says users of Fable 5.1 are far less likely to trip the safeguards that route requests to more restricted responses. Medical and biology questions will see 85 percent fewer interventions, and some users could see around 60 percent fewer cybersecurity-related interventions per session, according to the company’s statements reported by Axios.
This addresses a complaint developers have made loudly since June. Joseph Perla, founder of the startup Routing, told Axios that his users had been looking to reduce their reliance on Claude partly because of those interventions. Anthropic acknowledged at the time that it had launched conservatively and would loosen over time. This release is that loosening.
Privacy for enterprise buyers
Alongside the models, Anthropic introduced enterprise safeguards it says provide protections equivalent to zero-data-retention policies while still allowing the company to police malicious use. That is a direct answer to a sticking point in enterprise deals, where security teams have often refused to send confidential material to a provider that keeps logs. How the mechanism works in detail has not been fully described, but the pitch is clear: the privacy assurance of no retention, without giving up the ability to catch abuse.
The release as a playbook
Look at the whole package and a pattern emerges. New models, a cost cut aimed squarely at agent workloads, fewer refusals, tighter enterprise privacy, and a clear line between what is public and what is gated. Axios called it a possible new playbook for a launch: a smarter model is no longer enough on its own, it has to arrive with the tools and terms customers have been asking for.
Context helps explain the urgency. Anthropic is widely expected to file for an IPO this month, and shipping Fable 5.1 before OpenAI’s delayed Astra drew notice. Gavin Baker, managing partner at Atreides Management, called the timing “quite a flex” in a post on X.
The number to watch now is not a benchmark score. It is whether the 45 percent saving shows up in real agent deployments, where cache behaviour depends on how workflows are built. If it does, the calculation for running autonomous agents in production just got much easier, and every competing lab will need an answer on price, not just on capability. For more coverage of AI model releases and pricing, visit Mylistingo.
