Anthropic has put message threads, a speed feature for AI agents, into public beta, according to its Claude resources site.
Normally an agent working through a big task resends everything said so far with each new request, and a coding agent may fire off hundreds of them in a session. With message threads, the opening request carries the whole conversation, and later ones send just the fresh messages plus a pointer to the last response.
The payoff is a shorter pause before the reply starts. That pause is called time to first token: the gap between sending a request and the first piece of the answer appearing. Anthropic says requests of 500,000 to 1 million tokens see the median drop by about 35 percent, and most Claude Code users already have it on.
Cost does not move: Claude still reads the entire conversation each turn, so usage stays the same.
Anthropic keeps each thread encrypted in memory. If five minutes pass without a request, the thread lapses and the client starts over with a full resend. The 35 percent figure is Anthropic’s own.
Agents that stall past five minutes between steps lose the gain, so it favors busy sessions.
Reported by Anthropic on its Claude resources site on 2 November 2026.
