Claude Code just found a surprising speed boost in a tool created in October 1992. Anthropic says Claude Code now gzips the conversation it sends on every turn, helping each request reach the model faster and pushing local sessions to three times their August speed.
That change turns an old piece of compression technology into a fresh advantage for AI-assisted software development. Gzip was created by Jean-Loup Gailly and Mark Adler as a free alternative to Unix’s compress utility, and Claude Code now uses it across every turn of a conversation.
Claude Code’s Biggest Gains Arrive Locally
The improvement is not the same across every platform, but the gains are clear. Compared to August, Claude Code is now 3x faster on local sessions, 2.8x faster on cloud sessions, and 1.5x faster on API, chat, and Cowork platforms.
- Local Claude Code sessions are 3x faster than in August.
- Cloud Claude Code sessions are 2.8x faster than in August.
- API, chat, and Cowork platforms are 1.5x faster than in August.
Those figures focus on the wait before the first token, measured at the 90th percentile and excluding model processing time. That detail matters because it tracks the delay users feel before Claude begins responding, rather than the full time needed to complete a response.
The Claude API uses a different measurement. Its figure tracks the median time to the first token and includes model processing time, so the API number cannot be compared directly with the Claude Code wait measurement.
Anthropic also reports a separate improvement for Claude.ai: loading the service to a typeable page fell from 3.1 seconds to 0.55 seconds at the 75th percentile. Together, the figures point to a push that targets the entire path from request to response, not just the model itself.
Why Gzip Makes a Difference
Claude Code sends a conversation on every turn, and those conversations can contain the growing context of a coding task. Gzipping that data reduces what must travel before the model receives it, which helps shorten the wait for the first token.
The idea is simple, but its effect compounds across a working session. A developer who sends many turns does not experience one isolated transfer; each request carries the conversation forward. Compressing every turn gives the workflow a faster rhythm.
That practical focus matches the approach described by Anthropic’s Claude Code product lead, Cat Wu, who said, “We have no grand plan.” The phrase captures a product shaped by targeted improvements, including usage limits, transparency, and what Anthropic calls the “lean harness.”
Anthropic has raised Claude Code usage limits, and the company credits a new deal with SpaceX. Claude Managed Agents also gained higher limits, with usage limits doubled for Pro and Max users. Those changes give the speed improvements more room to matter because users can run more work before reaching a limit.
Claude Code Fits a Larger Agent Push
Claude Code is also moving beyond a traditional coding assistant. It can now take over a computer to complete tasks, with safeguards in research preview, while Anthropic launched Cowork, a Claude Code-like product for general computing.
Anthropic’s Claude Managed Agents can now “dream,” sort of. The phrase describes a new capability in a broader agent effort, where Claude handles work across longer sequences instead of answering one request at a time.
A Claude source leak revealed plans for a persistent agent, a stealth “Undercover” mode, and a virtual assistant named Buddy. These details sit alongside Anthropic’s pause on token-based billing for its Claude Agent SDK, showing how quickly the company’s agent plans are changing.
The wider Claude business is expanding at the same time. Google will invest as much as $40 billion in Anthropic, and Anthropic received a $5 billion investment from Amazon while securing 5 gigawatts of Amazon’s custom silicon. Anthropic could be worth $2 trillion when it goes public.
Anthropic is also preparing new model variants, including Opus 5.5, Sonnet 5.5, and Haiku 5.5. Opus 5 targets token efficiency instead of a capability leap, performing at the level of Claude Fable 5.1 on most tasks while costing 40% less to run than Opus 5. Sonnet 5.5 runs more than 30% faster and costs up to 30% less for most work.
The message is clear: AI performance is not only about producing smarter answers. It is also about reducing the pause before work begins, trimming the cost of each task, and giving agents enough room to keep going. A compression tool from October 1992 is helping Claude Code move toward that future, one conversation turn at a time.
Based on
