Anthropic Cut Claude's Cache Price 75%, From $1 to $0.25 a Million Tokens. Here's Where Your AI Bill Actually Drops
If you don't build on the Claude API, none of this touches your bill directly. Claude.ai's consumer subscription price hasn't moved. This is a change to what developers pay for, which matters to you exactly to the extent that some app you already use runs on Claude behind the scenes.
Prompt caching is the part of an AI bill that actually moves for a busy app. Instead of paying full price every time a chatbot re-reads the same long system prompt, or a coding agent re-reads the same codebase, the app pays once to write that content to a cache and a fraction of that price to read it again on the next request. Anthropic's own pricing page shows what that fraction now is on Fable 5.1: a quarter of what it cost on Fable 5, for the exact same $10-per-million-token model.
| Price component | Claude Fable 5 | Claude Fable 5.1 |
|---|---|---|
| Base input | $10 / MTok | $10 / MTok, unchanged |
| Output | $50 / MTok | $50 / MTok, unchanged |
| 5-minute cache write | $12.50 / MTok | $12.50 / MTok, unchanged |
| 1-hour cache write | $20 / MTok | $20 / MTok, unchanged |
| Cache read (hit) | $1 / MTok | $0.25 / MTok, a 75% cut |
75%
cut to the price of a cache hit on Fable 5.1 versus Fable 5
from $1 to $0.25 per million tokens, per Anthropic's pricing docs.
0.025x
Fable 5.1's cache-read multiplier, versus the standard 0.1x every other Claude model uses
3
breaking API changes shipped alongside the price cut
forced tool use, one-directional thinking blocks, and edit-invalidated reasoning, per Anthropic's own release notes.
1M tokens
Fable 5.1's context window, priced the same per token across the full window
The three breaking changes, in plain English
| What changed | What actually breaks |
|---|---|
| Forced tool use is gone | Setting `tool_choice` to force one specific tool now returns a 400 error instead of a response. Code that always forces a particular tool call has to switch to the default 'auto' mode, per Anthropic's release notes. |
| Thinking only flows one direction | Fable 5.1 can read an older model's reasoning, but no older Claude model can read Fable 5.1's. An app that routes a conversation to a different model mid-session silently loses that reasoning the moment it switches away from Fable 5.1. |
| Editing history invalidates reasoning | Change anything before a reasoning step, even a small system-prompt tweak, and the next request that replays it fails. Conversations built on Fable 5.1 have to be append-only, never rewriting an earlier turn. |
None of these are bugs. Anthropic's own documentation frames all three as consequences of Fable 5.1's reasoning being 'always on': the model thinks before most tool calls by default, and that reasoning now carries enough state that the API checks it hasn't been tampered with, per Anthropic's release notes. The cost is that a Fable 5 integration built casually, rebuilding the system prompt between requests, forcing one specific tool, switching models mid-session, breaks the moment it points at 5.1 instead of 5.
Who actually sees the savings
- 1Building on the Claude API with a long, reused system prompt or document context: switching the model ID to Fable 5.1 applies the cache-read savings automatically, no other code change required.
- 2Running a long agent loop that repeatedly re-reads the same context: this is where Anthropic's estimated savings of up to roughly 45% on heavily agentic workloads shows up, since those loops read from cache far more often than they write to it.
- 3Using Claude through an app you don't control, a writing tool, a customer-support bot: you won't see a price change directly. Whether that app passes the savings on is up to whoever built it.
- 4Migrating existing Fable 5 code: check for forced `tool_choice`, mid-conversation model switching, and any code that edits earlier messages, before flipping the model ID, per Anthropic's migration guide.
Did Claude get cheaper?
Only on the API. The base price per token is unchanged at $10 input / $50 output per million tokens. What dropped 75% is the price of reading from the prompt cache, from $1 to $0.25 per million tokens.
Does this change what I pay for a Claude.ai subscription?
No. This is API pricing for developers building on Claude. Claude.ai's consumer plans aren't affected.
Why would a 'thinking block' break my code?
Fable 5.1 reasons before most tool calls by default, and that reasoning is now tied to the exact conversation state that produced it. Editing an earlier message invalidates it, and older Claude models can't read it at all.
Who benefits most from the cheaper cache reads?
Apps and agents that repeatedly re-read the same long context, like a coding agent working through a large codebase or a chatbot with a long system prompt. Anthropic estimates savings of up to roughly 45% on heavily agentic workloads.
Is Claude Fable 5.1 a new model or an automatic update?
It's a new model ID, claude-fable-5-1, that developers have to actively switch to. It doesn't replace Fable 5 automatically in existing integrations.
Try the tools ✦
Free browser tools that never upload your files.