Editorial illustration for Claude Code Now Uses 34-Year-Old Gzip for 3x Faster Response
Claude Code Gets 3x Faster With 34-Year-Old Gzip
Anthropic shaved serious time off Claude Code's response speed, and the fix wasn't a new model or a server upgrade. It was gzip, the compression format Jean-Loup Gailly and Mark Adler released in October 1992 as a patent-free alternative to Unix's compress utility. Local sessions are now 3x faster than they were in August, cloud sessions run 2.8x faster, and the API, chat, and Cowork platforms all clock in at 1.5x. The numbers come from p90 measurements of wait time before the first token, stripping out model processing time, so the gains reflect improvements to transmission and session startup rather than the model itself doing less work.
The move fits a pattern. Last month Anthropic ran a two-week engineering sprint that merged more than 3,000 changes to speed up Claude's web and desktop interfaces, cutting the time to load a typeable page on Claude.ai from 3.1 seconds to 0.55 seconds at the 75th percentile. That effort landed around the same time Anthropic rolled out new Opus, Sonnet, and Haiku models, with Opus 5.5 aimed at agentic coding and professional knowledge work. Here's how Anthropic explained the latest change.
Gzip was created by Jean-Loup Gailly and Mark Adler as a free alternative to Unix's compress utility, whose LZW algorithm was subject to patents. First released in October 1992, gzip remains widely used to compress data for storage and transmission.
Why this matters
Anthropic's fix here is a reminder that not every speed gain in AI tooling comes from a bigger model or a smarter architecture. Gzip predates the web browser, and Anthropic reached for it anyway because the actual bottleneck was payload size on every conversational turn, not inference itself. For developers running Claude Code locally or in the cloud, a 3x drop in p90 latency is the kind of change you feel immediately in daily use, especially in tight edit-test loops where every second of lag compounds.
It's also a useful signal for founders building on top of Claude Code: performance complaints aimed at "the model" are sometimes really complaints about the plumbing around it. Worth watching is whether Anthropic publishes harder numbers beyond the X post, and whether these gains hold up outside the p90 benchmark once more users pile onto cloud sessions. Compression won't fix everything, but this is a cheap lesson in checking the unglamorous parts of the stack before assuming you need a faster model.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
From Archives Physics-Labeled Prompts Test LLM Calibration vs Four Ablations and Optimizers
From Archives Anthropic Gives AI Watchdog Employee-Like Access as Labs Propose Third-Party Evaluators
From Archives Anthropic to Allocate 5% of Stock for Public Benefit Fund
Business & Startups
From Archives Anthropic Nearly Stops Junior Hiring as Claude Agents Log Six Times Hours
Business & Startups
Common Questions Answered
Why did Anthropic use gzip compression to improve Claude Code's response speed?
Anthropic identified that the actual bottleneck in Claude Code's performance was payload size on every conversational turn, not inference itself. By implementing gzip compression, they were able to reduce the data being transmitted, resulting in significantly faster response times across all platforms without requiring a new model or server upgrade.
What specific performance improvements did gzip compression deliver for Claude Code?
Local sessions achieved 3x faster response times, cloud sessions reached 2.8x faster speeds, and the API, chat, and Cowork platforms all experienced 1.5x improvements. These measurements are based on p90 latency metrics for wait time before the first token, representing substantial gains in user experience.
When was gzip created and why was it developed as an alternative to Unix compress?
Gzip was released in October 1992 by Jean-Loup Gailly and Mark Adler as a patent-free alternative to Unix's compress utility, whose LZW algorithm was subject to patent restrictions. Despite being over 30 years old, gzip remains widely used today for compressing data for both storage and transmission purposes.
How does using an older compression tool like gzip challenge assumptions about AI performance optimization?
Anthropic's use of gzip demonstrates that not every speed improvement in AI tooling requires a bigger model or smarter architecture. This approach reminds developers that sometimes the most effective solutions come from identifying the true bottleneck—in this case payload size rather than inference speed—and applying proven, existing technologies to solve it.
Further Reading
- Claude Code is 3x faster to first token: what gzip and pre-started cloud sessions change - 7minAI News
- Claude Code waits 3x less for the first token: gzip on every turn - AICoder
- Claude Code cuts first-token latency threefold with gzip - Azeem Hassni
- Anthropic announces gzip for Claude Code and faster cloud sessions - Pasquale Pillitteri
- Claude Code Is 3x Faster to First Token: What Changed - CellCog
