Anthropic's message threads cut TTFT 35% on 1M-token agent sessions
ClaudeDevs · x · 2026-11-03
Anthropic shipped message threads in public beta: agents send only new messages each turn instead of the full history.
How it works
- First request carries the full conversation and starts a thread; later requests send only new messages plus the previous response ID
- Threads live encrypted in memory (never on disk), expire after 5 idle minutes
- Claude still reads the full context each turn, so token usage is unchanged
Results in Claude Code (on by default for most users)
- Median TTFT down 35% on 500K-1M token requests
- Median request size in long conversations: 800 KB → 1.8 KB
Availability: add the thread parameter and beta header; available on Claude Platform, AWS, Microsoft Foundry, and Google Cloud at standard token rates. Best for latency-sensitive active sessions: long-running coding agents, computer use (screenshots stay with the thread), document/image workloads.
Related event: Anthropic Launches Message Threads Beta for Messages API(4 posts)→
More from coding & agent
- Sonnet turns AGENTS.md update request into 'bless this repo' ASCII shrine — nptacek · 2026-11-03
- Developer vibe-codes his own trading terminal to ditch TradingView's attention traps — builderjaydub · 2026-11-03
- Chi Wang launches shared rooms for multiplayer agent collaboration, permission design flagged as key — anthara_ai · 2026-11-03
- Hackathon trick: intercepting Amazon Dash Button calls as free IoT code triggers — nptacek · 2026-11-03
- Paperclip Datasets launches: describe a table, agents gather sourced data from papers and patents — james_y_zou · 2026-11-03
- W&B adds linked zoom/pan for image logs; Karpathy built HackerTube in a weekend — wandb · 2026-11-03