Claude's Messages API adds threads: 35% faster first token

Beta on the Messages API: the server holds your conversation 5 min so you ship only new turns — token bill unchanged. Plus Claude Code's product lead exits.

Nowline NOV 4 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Send only the new turn — the server keeps the rest

    Message Threads, now in beta on the Messages API (`/v1/messages`), parks your conversation in encrypted server memory so each turn ships only the new messages — keyed off the prior response's `previous_message_id` — instead of replaying the whole transcript. Anthropic clocks ~35% lower median time-to-first-token on Claude Code requests carrying 500K–1M tokens of context, and it's already on by default for most Claude Code users.

  • The catch: same token bill, 5-minute memory

    Threads shrink upload size and latency, not cost — Claude still reads the full context every turn, so your token usage is identical (pair it with prompt caching to actually save). A thread expires 5 minutes after its last request and is purged within an hour; a lapsed one returns a 404, so keep your own copy of every message. No zero-data-retention or HIPAA, and no Batches, token counting, or on-demand compaction.

  • Claude Code loses its product lead

    Cat Wu, who ran product for Claude Code and Cowork for two years, announced on Nov 3 that she is leaving Anthropic; no successor or reason was named. It's another shuffle at the top of the tool many of you live in daily, weeks out from Anthropic's planned IPO.

  • Honeycomb looses 'AI SRE' agents on prod

    At its Nov 3–4 Innovation Week, Honeycomb pushed its Team Agents — 'AI SRE coworkers' that investigate incidents against live traces and high-cardinality telemetry and act only within limits you set — into early access. If you're shipping agents to production, it's a way to watch them (and your systems) without hand-rolling every dashboard.

  • Build this weekend: AWS's multi-agent docs assistant

    AWS published a blueprint for an agentic engineering-knowledge assistant — a planner routing to Catalog, Evaluation, and Knowledge specialist agents on Bedrock AgentCore with Nova Multimodal Embeddings, plus deterministic spec checks — shipped with notebooks and a repo to clone. A ready starting point for an internal docs agent over your own codebase and wikis.