Claude speeds up long conversations of AI agents

Anthropic launches Message Threads in public beta on Claude. The feature reduces messages sent to the API, with an announced 35% gain on long requests, without changing token consumption.

Anthropic launched Message Threads in public beta on November 2, 2026, on the Claude platform. This feature allows applications using the Messages API to send only new messages after an initial full request. The goal is simple: to reduce latency on long sessions, especially for code agents, agents using the computer, and workflows rich in documents or images.

Less data sent, not less context

With Message Threads, the application no longer sends the entire history at each turn. Instead, it transmits the ID of the previous response and the messages added since that response, while Claude reconstructs the conversation on the model side. Anthropic specifies that token consumption does not change, as the model still reads the entire exchange. The difference is therefore mainly in the round-trip time and the API validation work.

According to the documentation and figures cited by Anthropic, requests containing between 500,000 and 1 million tokens see the median time before the first token drop by about 35% when Message Threads is enabled by default for most Claude Code users. In other words, the text does not arrive faster because there is less of it, but because the infrastructure has less to redo. This nuance matters for uses where every second counts: a coding agent running for a long time, a tool navigating an interface, or a processing chain handling attachments.

A temporary memory, with fairly strict safeguards

Threads are kept in memory, encrypted and managed by the server, then deleted at the latest one hour after their last use. Their useful lifespan is even shorter: they expire five minutes after the last request, although each stored response restarts this timer. Anthropic also emphasizes a practical point: the application remains the source of truth and must keep its own copy of each message, as there is no API to list, read, or delete threads.

The mechanism is designed for active sessions, not for archiving. A thread contains the conversation sent to the model, the system and tools of the creation message, as well as a fingerprint of the fixed fields used to validate subsequent requests. Files, meanwhile, are kept by reference: the application must therefore keep all files accessible as long as the thread remains active. If continuation fails, the client must resend the entire conversation in a new creation request. On the streaming side, Anthropic asks to wait for the message_stop event before continuing; otherwise, a request too early may return a 404 error. Nothing very glamorous, but APIs have never been known for their sense of drama.

What the beta allows, and what it leaves aside

The beta is available on the native Claude platform, on Claude Platform via AWS, on Claude in Microsoft Foundry, and on Claude Models in Google Cloud, managed by Anthropic. It works with all models in beta but remains incompatible with several components: Message Batches, token counting, on-demand compaction, and most context modification tools. Threads are also not available for organizations that have enabled Inference hooks.

Another important limitation: Message Threads is not eligible for zero data retention nor HIPAA uses. However, Anthropic maintains that this feature does not affect other zero data retention options. For developers, the message is clear: the beta simplifies repetitive exchanges but does not erase compliance constraints nor the need to plan a fallback to classic mode if the thread disappears or a backup model takes over.

Key points

  • Public beta launched on November 2, 2026, on Claude.
  • Announced gain: 35% on 500,000 to 1 million tokens.
  • Threads expire five minutes after the last request.
  • Anthropic manages the beta on AWS, Microsoft Foundry, and Google Cloud.
  • Message Threads does not support Message Batches or token counting.
  • Claude Code enables the feature by default for most users.

By the numbers

  • 35% — median decrease in time before the first token, according to Anthropic.
  • 500,000 to 1 million tokens — size of requests highlighted for the gain.
  • 5 minutes — inactivity duration before thread expiration.
  • 1 hour — maximum delay before deletion of data in memory.

To read