Input remains $10 and output $50 per million tokens, while Anthropic estimates the cache discount can lower highly agentic workload costs by roughly 45%.
By RuntimeWire Staff · Published
Primary source: Forbes
Why it matters #
Persistent agents can reread the same instructions, repository state and documents across many turns. Fable 5.1's lower cache-read rate can reduce those bills substantially, but the result depends on cache reuse, prompt structure, output volume and how many turns the model needs to finish a task.
Anthropic released Claude Fable 5.1 and cut its cached-context read price from $1 to $0.25 per million tokens on September 1, a 75% reduction, while keeping input at $10 and output at $50 per million tokens. CEO and co-founder Dario Amodei is pricing repeated context below the equivalent charge for Claude Opus 5, a potentially meaningful difference for coding agents and other applications that carry the same material through long sessions.
Anthropic estimates the change will reduce costs by about 25% on typical workloads and by up to roughly 45% on highly agentic workloads. Those figures are Anthropic estimates and depend on cache-hit rates and session design.
Anthropic, a San Francisco-based public benefit corporation, was founded in 2021 by former OpenAI employees Dario Amodei, Daniela Amodei, Jared Kaplan, Sam McCandlish, Tom Brown, Jack Clark, Chris Olah and Ben Mann. Anthropic has built Claude into a family of consumer, enterprise and developer products while retaining its original focus on AI safety, interpretability and steerability.
The discount applies to one part of the invoice
An agent working through a repository, research assignment or spreadsheet can carry system instructions, tool definitions and earlier work through many turns. Prompt caching lets an application preserve a stable prefix and pay a reduced read rate when that material is reused. Anthropic's cost guidance says agentic tasks resend growing conversation context on every turn and recommends enabling prompt caching before other cost controls.
The practical effect depends on how much of an agent's billed input is fresh, how much is cached, and how much output the task requires. A workload with 100 million fresh input tokens and 900 million cached tokens would spend $1,900 on input under the old rates and $1,225 under the new ones, a 35.5% reduction in input cost.
The old bill consists of $1,000 for fresh input and $900 for cache reads. Under Fable 5.1 pricing, the cached portion falls to $225. Output charges remain unchanged and would reduce the percentage savings on the full task.
A short request composed mainly of new input receives little benefit. An agent that repeatedly reads a large, stable body of context can receive a much larger reduction. Edits to earlier context, changing tool definitions and poorly structured prompts can reduce cache reuse, so developers still need traces from their own applications to estimate the effect on a full invoice.
Fable 5.1 also occupies an unusual place in Anthropic's model documentation. Its standard input and output rates are twice those of Claude Opus 5: $10 and $50 per million tokens for Fable, compared with $5 and $25 for Opus. Cache reads reverse the comparison, with Fable priced at $0.25 per million tokens and Opus at $0.50.
That structure favors workloads where reused context accounts for a large portion of billed input. Total task cost still depends on how many turns a model needs, how much output it generates, latency and whether it completes the assignment reliably.
In Anthropic's customer section, Cognition co-founder and CPO Walden Yan said the company planned to move Opus 5 traffic in Devin to Claude Fable 5.1 on launch day, beginning with code review. Yan attributed the decision to lower cost per task. The statement is a customer endorsement published by Anthropic rather than an independent cost comparison.
Long-running assignments make caching matter
Anthropic positions Fable for coding sessions and knowledge work that can continue for hours, including managed agents operating across browsers, documents, spreadsheets, presentations and workplace applications. In those workflows, an agent can repeatedly ingest its instructions, accumulated state and reference material while taking actions across other software.
Lower cache-read prices can change the economics of that repeated input without guaranteeing a cheaper completed task. A model that loops unnecessarily, generates excessive output or loses a cacheable prompt prefix can erase part of the savings. An expensive model can still produce the lower final bill if it completes an assignment in fewer turns.
Dario Amodei trained as a biophysicist, worked at Google Brain and became OpenAI's vice president of research. His biography traces work on model scaling, safety and behavior, including research associated with GPT-2 and GPT-3. Anthropic President Daniela Amodei worked in global health and on Capitol Hill before joining Stripe as an early employee. She later managed engineering and people teams at OpenAI and became vice president of safety and policy. A Stanford Graduate School of Business profile says she wanted to work on technology with a clear social purpose.
Anthropic has ample capital to test whether lower cache prices produce more usage. Anthropic reported raising $30 billion in February at a $380 billion post-money valuation. It reported another $65 billion financing in May at a valuation of approximately $965 billion. Anthropic also said in its February announcement that Claude Code had passed $2.5 billion in annualized run-rate revenue; that figure was reported by Anthropic and was not presented as audited revenue.
Competitors are also building agents that run longer
OpenAI introduced the Codex app in February 2026 as an interface for supervising parallel coding agents over extended assignments. Google added built-in computer use to Gemini 3.5 Flash in June, allowing agents to act across browser, mobile and desktop environments. Cursor, a model-agnostic coding environment made by Anysphere, lets developers run agents with models from several providers.
This competition makes per-task economics more useful than a single token rate. A model with higher input prices can produce a lower bill if it completes the assignment in fewer turns or serves most repeated context from a cheap cache. A cheaper model can still cost more if it retries actions, generates excessive output or fails to preserve cacheable prompt prefixes.
The cache discount gives developers a measurable variable. Teams operating persistent agents can compare cache-hit rates, total task costs and completion quality before and after switching. Anthropic has left Fable 5.1's expensive standard token rates intact and reduced the charge for applications that keep reusable context stable.