Claude Code Speeds Up First Replies and Startup
Claude CodeClaude Code 2.1.283 shaved latency off both the start of a session and its first reply. A pattern-compile step that used to run only after a session's first reply finished now runs while that reply is still streaming, and first-request latency improved by reusing an already-open, preconnected API connection instead of opening a new one. Startup itself got lighter too: claude -p and Claude Code Remote no longer load the interactive terminal UI they do not need, and the auto-mode classifier's rules along with the Artifact tool now load on first use instead of at every launch. Accounts whose Artifact tool eligibility was not yet known, such as on a first run, no longer wait up to 1.5 seconds at startup to check it.
Key Takeaways
- The first reply of a session got faster, because a pattern-compile step that used to run after the reply finished now overlaps with the reply as it streams.
- First-request latency dropped through reusing an already-open, preconnected API connection instead of establishing a new one for the session's opening call.
claude -pand Remote sessions skip UI they never render, since both now bypass loading the full interactive terminal interface.- The auto-mode classifier and Artifact tool are now lazy-loaded, deferring their setup cost to first use rather than paying it at every launch.
- New or uncached accounts no longer stall at startup, since a check that could take up to 1.5 seconds to confirm Artifact tool eligibility now only blocks the first message that actually needs it.
- This continues a multi-week performance push, following earlier September releases that targeted CPU usage, fullscreen rendering, and large-session resume times.
Sources & Mentions
1 external resource covering this update
Faster First Replies
Claude Code runs a pattern-compile step early in a session that previously blocked completion of the first reply until it finished. Anthropic moved this step to run concurrently while the first reply streams to the user, so it no longer adds visible delay at a point in the session where users are watching for the first response.
Faster First Requests and Lighter Startup
First-request latency also improved by reusing an API connection that was already preconnected, avoiding the overhead of opening a new one for the first call of a session. On the startup side, Claude Code trimmed what loads before a session is ready: claude -p (the non-interactive, scriptable mode) and Claude Code Remote no longer load the full interactive terminal UI, since neither needs it, and the auto-mode classifier's rule set along with the Artifact tool now load lazily on first use rather than unconditionally at launch. For claude.ai accounts whose Artifact tool feature eligibility is not yet cached, such as on a brand-new account, startup previously could wait up to 1.5 seconds to check that eligibility before proceeding; now the first message waits for that check only if it is actually needed, rather than the whole session start doing so.