Anthropic Cuts Long-Context Agent Startup Delays
Anthropic’s message-thread API sends only new context on each turn, cutting median first-token latency by about 35% in huge requests.
Anthropic has introduced message threads for Claude-powered applications, a platform change designed to reduce the repeated transfer of long conversation histories. With the feature enabled, an application creates a thread once and sends only the new material on subsequent turns. Anthropic temporarily stores the thread server-side, but the application remains the source of truth and must retain its own conversation copy.
Anthropic says the change is especially useful for long-running agents and Claude Code workflows. For requests containing roughly 500,000 to one million tokens, the company reports a median reduction of about 35% in time to first token. Message threads are available in beta on Anthropic’s native platform and through its managed Claude offerings on AWS, Microsoft Foundry, and Google Cloud.
The improvement is primarily an infrastructure optimization rather than a new model capability. It does not make a model reason better or expand its formal context limit. Instead, it reduces the amount of repeated input that must be transmitted and processed for every turn. That distinction matters for agents, which may repeatedly inspect files, tool results, and prior decisions over hours of work.
The trade-offs are operational. Thread state becomes part of the application’s dependency on Anthropic’s platform, but threads are best-effort: they expire five minutes after the last request and are deleted no more than one hour after last use. Developers must also understand which parameters remain fixed across a thread. The documentation specifies restrictions around model, tool, thinking, context-management, and other configuration changes. Teams building portable multi-provider systems may therefore need an additional abstraction layer.
Why it matters: long-context agents are often limited by waiting, bandwidth, and token-processing overhead rather than by the model’s raw intelligence. A 35% latency improvement at million-token scale can make iterative coding and research workflows feel substantially more responsive, although Anthropic has not published a full cost comparison or independent benchmark.