Anthropic has published a detailed engineering account of how it made claude.ai and the Claude desktop app roughly three times faster in a two-week sprint, and the twist is that most of the work was done by Claude itself. The post, published on Anthropic's engineering blog, describes a performance campaign run from a single Slack channel with the AI model in every thread, finding bottlenecks, building benchmarks, shipping improvements, and watching every deploy.
The numbers, measured at the 75th percentile, are substantial. Time to a typeable page on a fresh load of claude.ai fell from 3.1 seconds to 0.55 seconds. Starting a new Claude Code session dropped from 0.8 seconds to 0.3 seconds, and loading a Claude Cowork cloud session fell from 2.6 seconds to 0.73 seconds. In aggregate, Anthropic estimates the improvements save tens of thousands of user-hours of waiting every day. Para mais contexto sobre esta história, confira nossa notícias de IA.
Measuring first, then moving
The sprint began with measurement rather than code. Using Claude to analyze usage data through a Datadog MCP server, the team identified the four user journeys that account for 95 percent of activity: launching the app, starting a conversation, loading an existing conversation, and sending a message. Across web and desktop those journeys broke down into thirteen distinct measurements, and the team added instrumentation until each was directly comparable — starting at a user interaction and ending once the result rendered.
With baselines in place, the team drew up a list of roughly twenty hand-picked projects, each targeting a specific journey, and had Claude estimate the impact of each in milliseconds to set sprint targets. The plan landed early: Anthropic reports hitting twelve of the thirteen targets by day three. That left room for Claude to identify further opportunities and propose new workstreams, which quickly ramped into full projects of their own and pushed results beyond the original goals.
What was actually shipped
The technical changes read like a checklist of modern web performance work. For faster launches, the team baked a static composer into the HTML so users can start typing during React initialization, and precompiled a V8 code cache so the desktop shell's main process no longer recompiles from scratch on startup. For faster navigation between conversations, the composer stays mounted rather than being torn down and rebuilt, sessions are prefetched when a user hovers over them, and sidebar re-renders were cut by 90 percent.
The scale of the merge history is the headline figure: more than three thousand changes landed during the sprint without a single customer-facing incident or rollback, according to the post.
Claude Tag: an agent with guardrails
The sprint ran on what Anthropic calls Claude Tag, currently in beta and described as an internal research model roughly comparable to Opus 5.5. The division of labor is the most consequential part of the story. Claude found bottlenecks, built benchmarks, implemented fixes, and monitored every deployment — working asynchronously for hours at a time, even overnight. Humans set the goals, made the tradeoffs, and approved every change before it merged.
The team also wanted to iterate faster than its deploy cadence, so it built lab measurements as proxies for real-world reads. Among the questions the engineers asked: could JavaScript instruction counts replace wall-clock timing as a faster feedback signal for validating prototypes before deployment?
Why it matters for AI-assisted engineering
Anthropic's post is notable less for the specific optimizations — static shells, prefetching, and render reduction are established techniques — than for what it demonstrates about AI-driven development at production scale. A frontier lab just shipped a three-thousand-change performance overhaul of a flagship consumer product with AI agents doing the bulk of the identification, implementation, and verification work, while humans retained approval authority over every change.
The result also lands in a competitive context where responsiveness is a product feature. Claude competes directly with OpenAI's ChatGPT and Google's Gemini for consumer and developer attention, and perceived speed shapes daily usage of assistant products as much as model quality does. Cutting time-to-interactive from 3.1 seconds to 0.55 seconds addresses the single most common complaint users had about the product.
Para as equipes de engenharia que observam de fora, a lição transferível do relato da Anthropic é a estrutura do loop: medir com precisão as jornadas do usuário, deixar o modelo propor e validar mudanças em relação a essas medições, manter uma porta de aprovação humana e expandir o escopo apenas na velocidade que o sistema de medição puder verificar. O próprio enquadramento da Antrópico é que, uma vez que Claude consegue medir algo, ele pode torná-lo mais rápido – então a equipe continuou encontrando mais coisas para medir.
Ainda não se sabe se sprints semelhantes conduzidos por agentes se tornarão uma prática padrão em toda a indústria de software, mas a Anthropic forneceu um dos pontos de dados públicos mais concretos que eles podem trabalhar em escala. Os leitores que acompanham como a IA está remodelando o desenvolvimento de software podem acompanhar nossa cobertura de pesquisa em IA para desenvolvimentos futuros.
---
Fique à Frente da IAAs últimas notícias, análises e avanços em inteligência artificial — tudo em um só lugar.
Ler mais notícias de IA →
