September 5, 2026. Anthropic released Claude Fable 5.1 on September 1 and left the headline API price where it was, $10 per million input tokens and $50 per million output. The change that matters to anyone running an agent is one line further down the announcement: cache reads now cost $0.25 per million tokens, 75% below Fable 5's $1.00. Anthropic puts the effect at roughly 25% off a typical workload and up to about 45% off highly agentic work. The same release quietly introduces a rule that applies to new API accounts before it applies to anyone else, which means a business starting its first agent this month is the first to feel it.
What changed, in numbers
- Cache reads fall from $1.00 to $0.25 per million. VentureBeat notes this is 2.5% of the input price against a 10% multiplier on other Claude models, so Fable 5.1's cached input is now half of Opus 5's ($0.50) and close to Sonnet 5's ($0.20).
- Everything else holds. Input $10, output $50, five-minute cache writes $12.50 per million and one-hour writes $20, batch processing at half price for asynchronous jobs.
- Anthropic's own estimate: about 25% lower cost on typical workloads and up to about 45% on context-heavy, tool-heavy agent work where cache reads are most of the bill.
- Preserved thinking, for new accounts first. Per Anthropic's help center, API organizations created on or after August 31, 2026 cannot edit the system prompt, tools or messages that preceded a thinking block in a multi-turn conversation. The API returns an error, or in an opt-in non-strict mode drops the affected thinking blocks. Existing accounts are exempt on Fable 5.1. Future models apply it to everyone.
- Enterprise Frontier Safeguards will keep monitoring data in the customer's own cloud, rolling out in phases this fall, with zero data retention for eligible customers until then.
What it means for operators
Work the cache number on a concrete agent. Say a support or lead-qualification agent carries a 60,000-token system prompt, tool definitions and knowledge base, and handles 200 conversations a day. That is 12 million cached tokens a day. At $1.00 per million it cost $12 a day; at $0.25 it costs $3. That is about $270 a month saved on one agent, before fresh input and output, which are unchanged. The saving only shows up if the prompt is actually cached, meaning a stable prefix that does not change between calls. Agents that rebuild their system prompt every turn get none of it.
The output side has not moved, and output is five times the input price. An agent that writes long emails or reports for every task is output-heavy and will see far less than 25%. The 45% figure belongs to agents that read a lot and write a little: research, monitoring, classification, routing. Know which kind you run before budgeting the discount, and read the GPT-6 Astra pricing alongside it, because OpenAI's flagship now sits at the same $10 and $50.
Then the rule that bites newcomers. Preserved thinking targets industrial distillation, and Anthropic starts with new accounts because that is where the abuse concentrates. The side effect is that a company whose first Claude Platform organization was created on August 31 or later is on the strict path from day one, while an account opened in July is not. The integrations Anthropic names as needing adjustment are exactly the ones a home-built agent harness tends to have: context compaction, which rewrites earlier turns to save tokens, and injected system reminders. If your automation does either, stop rewriting prior turns or opt into non-strict mode and accept that the model loses its earlier reasoning on those calls. The consolation Anthropic notes: unchanged thinking blocks let the API reuse cached prompts more often, which feeds the discount above.
VentureBeat also cites a Financial Times analysis of Ramp data across roughly 70,000 companies showing Fable 5 at about 11% of Anthropic spend two months after launch, with cheaper Opus models gaining share. A 75% cut on the token type that dominates agent bills reads as a direct answer to that. For a business choosing the model behind its agents, the question is no longer list price but cost per completed task on your own token mix, a measurement worth having an engineer run before the next invoice rather than after.
Frequently Asked Questions
The list prices are identical at $10 per million input tokens and $50 per million output. The change is cache reads, which fall from $1.00 to $0.25 per million, a 75% cut. Anthropic estimates this lowers the cost of typical workloads by about 25% and highly agentic workloads by up to about 45%, based on four weeks of August 2026 usage.
A cache read is the model re-reading context it has already processed and stored, such as a long system prompt, tool definitions or documents that are identical on every call. Agents repeat that context on every turn, so on tool-heavy, context-heavy work cache reads make up most of the tokens billed. That is why the discount is largest there.
On Fable 5.1, API organizations created on or after August 31, 2026 cannot modify the system prompt, tools or messages that came before a thinking block in a multi-turn conversation. The API returns an error, or in non-strict mode drops the affected thinking blocks. Existing accounts are exempt on this model. Anthropic says it will apply to all accounts for future models. Claude Code, Cowork, Claude.ai and third-party products are not affected.
No. Output tokens still cost $50 per million, five times the input rate, so output-heavy agents that write long documents see far less than the headline saving. The discount favors agents that read a lot and write a little. The useful comparison is cost per completed task on your own mix of cached, fresh and output tokens, not the list price.