Claude Fable 5.2 changes how custom agents reuse history
Anthropic now checks saved reasoning against earlier messages and tools on every Fable 5.2 account. A changing tool list can reduce reuse.
Anthropic's November 9 release of Claude Fable 5.2 changes what an application can carry forward in a conversation. The API now checks saved reasoning against the messages, instructions and tools that were in place when Claude produced it. That default applies to every Fable 5.2 account. On Fable 5.1 it applied by default to accounts opened on or after August 31, 2026.
When a tool list changes
An agent might use a connector to look up inventory or pull a report. If an app uses Anthropic's MCP connector and lets the tool list update automatically, a change mid-session can invalidate earlier reasoning on Fable 5.2. Anthropic's migration guide says that particular change did not affect saved reasoning on Fable 5.1 or Opus 5.5. When a custom client resends affected reasoning on a standard Messages request, the default is to reject it. Sending the same request again leaves the conflict in place.
This matters most when an application assembles and resends its own conversation history. Anthropic says its services that manage the history, including Claude Code, Claude Managed Agents and the Agent SDK, already handle this particular change.
Cheaper tokens and the cost of finishing a task
Anthropic cut standard list prices by 20% across the entries below. A documented continuation option can discard affected reasoning so the request can proceed. The cache restarts at the edit, and Claude may need to generate more output to reconstruct work it could otherwise reuse.
| USD per million tokens | Fable 5.1 | Fable 5.2 |
|---|---|---|
| Uncached input | $10.00 | $8.00 |
| Output | $50.00 | $40.00 |
| Cache reads | $0.25 | $0.20 |
| 5-minute cache writes | $12.50 | $10.00 |
| 1-hour cache writes | $20.00 | $16.00 |
Anthropic's standard API price list. A cache read uses previously saved input. Cache writes store input for later requests.
At the same mix of billed token types, 25% more volume would erase a 20% price cut. A change in the mix matters too. Replacing cached input with newly processed input can move tokens into a more expensive category. The relevant comparison for a long-running agent is its bill for a completed task, including how often it can reuse earlier work.