AI TrackerGet the app
← All news

SemiAnalysis: long contexts add little latency on Claude Opus 5.5 when cached, but cold starts cost

SemiAnalysis measured time to first token (TTFT) from its own internal use of Claude Opus 5.5: 632K requests run between October 6 and November 8. These are the firm's figures from its own traffic, not a vendor benchmark.

  • With a cached context, p50 TTFT rose from 0.9s at 50–100K tokens to 1.6s at 600K+ tokens.
  • At 600K+ tokens, p50 TTFT was 3.5s uncached against 1.6s cached, about 2x. At p90 the gap was 3.5x: 10.7s uncached against 3.1s cached. "Uncached" means more than 50K tokens had to be processed from scratch.
  • Only 1.8% of requests were that uncached, but re-processing their context made up about 15% of its Opus 5.5 spend. Most were new subagents starting from scratch, and 12% of sessions paid about $3 to resume after more than an hour idle.

The thread had 29 points and 15 comments on X on November 17.

Get news like this as a push.

AI Tracker alerts you when something changes for the tools you use.

Get the app