Claude Code cuts first-token latency ~3x, to 1.33s locally
Gzipped turns and cloud sessions that warm as you type drive the drop — cloud still waits ~6s. Plus: Claude traced the NIST curve seeds to a 1990s clock.
Copy markdown
Local: 4.07s → 1.33s first token
Anthropic cut p90 time-to-first-token between August and October: local Claude Code 4.07s→1.33s, cloud sessions 16.99s→6s, the API 1.53s→1.01s, and Claude web/Cowork 3.36s→2.17s. If you live in the terminal, the slowest part of the loop — waiting for the model to start — is now about a third of what it was.
The trick: gzip the turn, warm the cloud early
Two changes do most of the work: large requests to api.anthropic.com are now gzip-compressed (biggest win on long contexts, shipped in Claude Code v2.1.300 on Oct 14), and cloud sessions begin warming the moment you start typing instead of after you hit enter. Nothing to configure — just update Claude Code and the speedup is on.
The off switch and the fine print
You can disable request compression with CLAUDE_CODE_GZIP_REQUEST_BODIES=0 if it ever misbehaves. Caveats worth knowing: gzip is hardly new, pre-warming hides rather than removes cloud startup (so a ~6s cold wait is still noticeable), and spinning up compute on each keystroke means prompts you never send can still cost a start.
Research radar: Claude cracked the NIST curve-seed riddle
Anthropic says Claude traced the unexplained seeds behind the NIST/NSA ECDSA curves — a ~26-year open question — to the lrand48 PRNG seeded with Unix timestamps, in roughly 30 minutes of agentic work. Not a workflow change, but a sharp data point on what a coding agent can do in deep, unfamiliar territory.