On 1 September 2026, Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1, one model under two levels of safeguards, with a 75% cut in the price of cache reads (VentureBeat). For a company already using Claude, the real saving sits between 25 and 45% according to Anthropic, and depends on how its teams work with the tool: the price of input and output tokens does not move.
The facts
- The price of cache reads falls from $1 to $0.25 per million tokens, with no change to input ($10) and output ($50) rates. Cache writes stay at $12.50 for five minutes and $20 for one hour; by comparison, Opus 5 reads its cache at $0.50. The model is available on the Claude API, AWS, Google Cloud and Microsoft Azure (VentureBeat, Vellum).
- On Terminal-Bench-Science 0.1, which evaluates a scientific research task carried out end to end in a terminal, Fable 5.1 reaches 52.6%, against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol in Anthropic's configuration, with a margin of error of 3.5 to 4.5 points per model. On Terminal-Bench 4.0, the score is 55.8%, and 60.9% for Mythos 5.1. These are the vendor's own results (Vellum).
- Anthropic reports 60% fewer false positives on its cybersecurity safeguards and 85% fewer triggers on benign medical and biological questions. Fable 5.1 can now be used to discover software vulnerabilities, without developing exploits; life sciences research goes through Mythos 5.1 and a verification programme built with the US government (Anthropic).
- Cognition, the company behind the Devin coding agent, integrated Fable 5.1 on launch day and published its own calculation: on a FrontierCode 1.1 Extended task, the cost drops from $5.84 with Fable 5 to $2.68, below the $3.51 of Opus 5, because more than 95% of the tokens in an agentic task are cache re-reads (Devin).
Why does this price cut benefit some uses of Claude far more than others?
The cut only applies to cache reads, meaning the context Claude has already processed and reads again from one request to the next: the code repository, system instructions, tool definitions, attached documents, conversation history. Anthropic puts the saving at around 25% for typical usage and up to 45% for heavily agentic workloads, where an agent re-reads a large context before each action. Cognition's calculation shows the mechanics: a typical Devin task reads around 3 million tokens from cache, against 70,000 tokens of new input and 21,000 output tokens. At the previous cache rate, re-reading accounted for more than 60% of the bill; at the new rate, it accounts for less than 30%.
At the other end of the spectrum, a team that asks short questions in the Claude interface, with no project, no attached document and no connector, reuses little context and will see its bill change only at the margin. The list price remains the highest in the range: Opus 5 costs half as much for input and output. The right question is therefore how the teams actually work with the tool. Cognition also notes that Fable 5.1 completes the same tasks with 33% fewer tokens than Opus 5, a second source of savings that again depends on the type of work assigned.
Fable 5.1 and Mythos 5.1 remain one and the same model. Fable is publicly available, with Anthropic's production safeguards; Mythos, with lighter safeguards, is reserved for verified organisations in defensive cybersecurity and life sciences, for now in the United States. The launch closes a turbulent sequence: in June, the US administration ordered Anthropic to cut off access to Fable 5 and Mythos 5 for all foreign nationals, which led the company to disable the models worldwide, before the Department of Commerce lifted its export controls and access was restored from 1 July (Al Jazeera). For a European company, the episode is a reminder that a model's availability is also a political decision.
AIxH's view
Each new version of Claude moves the line between the tool you test and the tool you put into production, and few teams take the time to recalibrate it. That is the purpose of our Claude training in Luxembourg, built on participants' real use cases: projects and memory, MCP connectors, Cowork, and now the choice between Fable 5.1, Opus 5 and Sonnet 5 according to how much context each task reuses. The drop in unjustified refusals matters just as much in the training room: cases participants had set aside because the tool refused to answer become workable again. This module is part of the AIxH AI training programme, delivered by an approved continuing vocational training organisation in Luxembourg. If your teams still work with the Claude reflexes of a year ago, a one-day workshop is enough to bring them up to date.