Claude message threads: replay, retention and API costs
Anthropic introduced message threads beta on November 2, 2026. Later requests send only new content. Enable message-threads-2026-08-12. Access covers Claude API, Claude Platform on AWS, Anthropic-managed Google Cloud and Microsoft Foundry.
| Step | Integration requirement |
|---|---|
| Start | Submit complete history with thread: {"type":"create"}. |
| Continue | Send new messages and the latest response ID. Omit system/tools. |
| Recover | Keep local history. Recreate after a missing-thread 404. |
| Stream | Wait for message_stop before continuing. |
Each response allows one continuation. Threads expire after five idle minutes. Stored responses restart that window. Deletion follows within one hour of last use. Threads are not ZDR eligible and exclude HIPAA-ready organizations. Batches and token counting are unsupported. Pricing and rate limits stay unchanged. Message threads guide, November 2, 2026 release notes.
Caching needs its own check
Prompt caching is a separate opt-in for Messages API requests. Its own guide describes automatic top-level configuration or explicit block-level breakpoints. A cache stores reusable prompt prefixes, rather than serving as the application’s history database.
For Sonnet 5.5, the documented minimum cacheable prompt is 512 tokens. A shorter prompt can run normally without creating a cache, even when marked for caching. Check the response’s cache-creation and cache-read token counters rather than assuming that a successful request means a cache hit. If both counters are zero, the prompt was not cached. Prompt caching documentation.
Price the workload, not just the request size
These are current standard USD rates per million tokens, checked November 2, 2026, not a separate thread subscription:
| Model | Input | Output | Five-minute cache write | Cache hit |
|---|---|---|---|---|
| Sonnet 5.5 | $2 | $10 | $2.50 | $0.10 |
| Opus 5.5 | $4 | $20 | $5 | $0.20 |
For an illustrative Sonnet calculation, 10,000 ordinary input tokens cost $0.02. The same quantity written into a five-minute cache costs $0.025. A later cache read costs $0.001. Those figures exclude output and other charges. Reuse frequency matters: the first cache write is not itself a discount. Current pricing.
Before rollout
Treat adoption as an integration change, not a benchmark result. Exercise recovery in a development environment, compare observed token accounting with your cost estimate, and review whether your data-handling requirements permit this feature. ByteForward has not run hands-on latency or reliability tests for this report.
AI-generated conceptual illustration by ByteForward, not a product interface.
