Anthropic said on November 2 that message threads are now in public beta on the Claude Platform, letting long-running agents send only what is new on each turn instead of resending the full conversation every time. The change targets the growing cost of agent sessions, where each request has carried every earlier message and tool result, and it is already switched on by default for most Claude Code users.
How message threads works
Without a thread, an agent that makes hundreds of requests in one session uploads the entire conversation again on every turn. The API then validates and tokenizes all of it before Claude can start responding, and that repeated work grows with context, so the longest sessions wait longest before the first token arrives.
Anthropic reports that the saving is real where context is largest. For requests carrying 500,000 to 1 million tokens, the median time to first token fell by about 35 percent with message threads enabled for most Claude Code users. In long conversations, routine requests shrank from roughly 800 kilobytes to about 1.8 kilobytes.
Developers switch the feature on by adding a thread parameter to Messages API requests and using the designated beta header. The first request sends the full conversation and starts the thread, and each later request sends only the new messages plus the ID of the previous response, which becomes the key for the next turn.
What it does not change
Token use stays exactly the same, and Anthropic is explicit about why. Claude still reads the full conversation on every turn, so input-token charges and existing rate limits are unchanged. What shrinks is the size of the request travelling to the API and the time spent validating it.
Threads are deliberately short-lived and best-effort, not a new memory store. A thread expires five minutes after its last request, and its contents are held in encrypted memory, never written to durable storage, and deleted no more than one hour after last use. Applications must keep their own copy of every message.
Limits developers must design around
The beta carries several hard limits that shape how it can be used. Threads are append-only, each response ID can be continued only once, and changing a create-only field such as the system prompt or tools requires a fresh thread with the full conversation included again.
There is also no API to list, read, or delete threads, and the API does not store turns written by fallback models. A fallback therefore forces a full resend before threading can resume. Threads are not eligible for zero data retention or HIPAA use, and they are unavailable to organizations with Inference hooks enabled.
Our read
The honest framing is that this is plumbing, not magic, and that is why it matters. Agent frameworks have spent the last year fighting context bloat with compaction and retrieval tricks, all of which risk losing detail. Message threads attacks a dumber problem instead: the repeated upload of context Claude already holds.
The trade-off to watch is the five-minute expiry and the best-effort promise. A coding agent that pauses while a human reviews a plan will hit expired threads often, and each expiry means a full resend anyway. Anthropic’s warning that applications remain the source of truth is the right one for builders.
For Anthropic, the timing is strategic as well as technical. The company has recently warned that AI-assisted development may soon accelerate faster than its ability to evaluate it, as set out in its amended filing language on self-improving AI. Faster agent loops make that acceleration more likely, so the guardrails around threads matter too.
Why It Matters
Long-running agents are where the frontier labs now compete, and latency is the tax every agent pays on every turn. A coding agent that waits longer before its first token on each of hundreds of requests feels sluggish even when the model itself is fast, and that friction shapes which tools developers keep using.
The beta is also available broadly from day one: on the native Claude Platform, on Claude Platform on AWS, in Microsoft Foundry, and on Claude Models on Google Cloud under Anthropic management, with beta support on every model. That breadth puts pressure on rivals to match the same transport efficiency.
Background
Message threads arrives weeks after Anthropic’s amended public offering filing disclosed hundreds of billions of dollars in infrastructure commitments, a reminder that the agent era is expensive on every axis. The beta does not cut that bill, but it cuts one of its most annoying frictions.
Source: Anthropic announcement and Unite.AI report
0 Comments