Anthropic Adds On-Demand Conversation Compaction to Claude API
A beta returns a signed summary block on a developer-chosen request, letting applications replace older history while retaining selected recent turns.
Edited by Tyronne Panaino
Anthropic added on-demand conversation compaction to the Claude Messages API on September 14, 2026. The beta gives developers a way to ask for a summary at a chosen point in a long-running conversation, then use that result in later requests instead of resending all of the messages it represents.
The change matters most to teams building agents and assistants that may accumulate lengthy histories. It turns compaction into an application-controlled step that can run in the background, while the live conversation continues on its full history until the application is ready to substitute the summary.
What the API now returns
Anthropic's release note says developers enable the feature with the `compact-2026-09-04` beta header and send a top-level `compaction` parameter. The API returns a signed `compaction` block that summarizes the submitted messages. A later request places that block first, replacing the messages it covers.
Applications can keep recent turns word for word after the summary rather than compressing the entire conversation into one opaque checkpoint. Anthropic also says preserved thinking in retained turns can remain valid on models that support that behavior. The announcement describes the mechanism, not a guarantee that every detail in a long history will survive summarization.
Why the control matters
The practical delta is timing and ownership. A developer can decide when a history should be compacted instead of waiting for the next ordinary response to perform the maintenance work. Because the compaction request can run in the background, an application can prepare a smaller context without immediately interrupting the active exchange.
That creates a clearer engineering boundary for long-running work: the service produces a signed summary block, while the application decides when to replace older messages and which recent turns to keep. It may simplify context-window management for persistent agents, but teams still need to test whether their own instructions, tool state and important decisions remain usable after the swap.
Evidence limits and next checks
The evidence is one official Anthropic release note, so internal confidence is medium. The note establishes the beta, header, response shape and continuation pattern. It does not provide independent measurements of summary fidelity, latency, token savings or performance across different workloads.
The next useful checkpoints are broader availability, detailed failure reporting and reproducible tests that compare an agent's behavior before and after compaction. Developers should also verify that their audit trail retains the original messages when their governance or debugging process requires them.
Status
Confirmed. Anthropic documents the on-demand compaction control as a Claude API beta released on September 14, 2026.
Sources
Update note: Last reviewed 2026-09-17. We will revise this post when Anthropic changes availability or publishes additional implementation evidence.
Sources
- Anthropic — Claude Platform release notes — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.