Anthropic Introduces Claude Message Threads to Reduce API Payload Size and Latency
Anthropic has launched Claude Message Threads, a new feature in its Messages API that reduces conversational payload size and latency for multi-turn and agentic AI workflows. Instead of resending entire dialogue histories with each request, applications now send only incremental messages plus the previous response identifier, with the server persisting conversation state. Anthropic reports this cuts median payload size from roughly 800KB to 1.8KB in long dialogues and reduces time-to-first-token latency by about 35%. The change does not alter token-based billing, rate limits, or pricing, since the underlying model still processes the full context on each pass. This is a product/performance announcement with no security vulnerability or incident implications.
Detalhes técnicos
Threads are created with an initial request containing the full dialogue history; subsequent requests submit only the newest prompt and the prior message ID. Threads automatically expire after five minutes of inactivity, after which referencing an expired thread returns an error, requiring clients to retain local transcripts to reconstruct sessions. Threads support only sequential append-only operations, and core parameters like system prompts and tool schemas must remain constant for the thread's lifetime; any changes require a new thread. The feature is designed to complement prompt caching, combining reduced network overhead with discounted cache-read compute costs for eligible context segments.
Linha do tempo
-
Anthropic Introduces Claude Message Threads to Reduce API Payload Size and Latency
ler a análise desta fonte ocultar
Anthropic has launched Claude Message Threads, a new feature in its Messages API that reduces conversational payload size and latency for multi-turn and agentic AI workflows. Instead of resending entire dialogue histories with each request, applications now send only incremental messages plus the previous response identifier, with the server persisting conversation state. Anthropic reports this cuts median payload size from roughly 800KB to 1.8KB in long dialogues and reduces time-to-first-token latency by about 35%. The change does not alter token-based billing, rate limits, or pricing, since the underlying model still processes the full context on each pass. This is a product/performance announcement with no security vulnerability or incident implications.
Technical Details
Threads are created with an initial request containing the full dialogue history; subsequent requests submit only the newest prompt and the prior message ID. Threads automatically expire after five minutes of inactivity, after which referencing an expired thread returns an error, requiring clients to retain local transcripts to reconstruct sessions. Threads support only sequential append-only operations, and core parameters like system prompts and tool schemas must remain constant for the thread's lifetime; any changes require a new thread. The feature is designed to complement prompt caching, combining reduced network overhead with discounted cache-read compute costs for eligible context segments.