Anthropic makes Claude.ai 3x faster using Claude

The 3x Speedup

Anthropic has significantly accelerated the core user experience of claude.ai and its desktop application, achieving a roughly threefold speedup in a two-week sprint. According to the official Anthropic engineering blog, the company focused on four specific user journeys that account for 95% of user activity. The results are substantial: time to a typeable page on a fresh load dropped from 3.1 seconds to 0.55 seconds, starting a new Claude Code session improved from 0.8 seconds to 0.3 seconds, and loading a Claude Cowork cloud session decreased from 2.6 seconds to 0.73 seconds. In aggregate, Anthropic estimates these optimizations save tens of thousands of user-hours of waiting time every day.

The project was executed with a high degree of automation, utilizing Claude Tag (beta) running an internal research model comparable to Opus 5.5. The model was tasked with finding bottlenecks, building benchmarks, shipping improvements, and monitoring every deployment. Human engineers steered the process by setting goals, making tradeoffs, and approving every change. This collaborative approach resulted in more than 3,000 merged changes without a single customer-facing incident or rollback.

Measuring the Unmeasurable

The core challenge in performance optimization is often the lack of precise, deterministic metrics. Traditional wall-clock timing is noisy and difficult to use as a continuous integration (CI) gate. To solve this, the team moved away from simple latency measurements and toward deterministic counts. They utilized Valgrind with node --predictable to count JavaScript instructions for pure-JS hot paths. For browser-based paths, they tracked React commits, function call counts from V8’s precise coverage, layout and style-recalculation counts, and DOM mutations.

Each new benchmark served two purposes: it provided a metric Claude could optimize in the lab, and it acted as a guardrail in CI with a number that could only ratchet down. If a benchmark was flaky or did not correlate with actual user latency, it was discarded. This rigorous validation process ensured that the model was optimizing the right variables. For instance, when optimizing the routine that assembles a conversation’s message tree, Claude profiled the path with Valgrind and discovered that a quarter of the instructions were megamorphic dictionary lookups resolving the same message ID three separate times. After refactoring, instruction counts dropped by 48% and 31% on two hot paths, correlating with wall-clock time reductions of 78% and 44%, respectively.

The Autonomous Loop

The workflow was centered around a single Slack channel where Claude operated with standing instructions to facilitate all performance-related tasks. Its responsibilities included monitoring deploys for regressions, assessing telemetry accuracy, maintaining observability dashboards, and proactively implementing solutions. The team encouraged autonomy, with instructions stating, "The ultimate goal for this channel is for you to become as autonomous as possible, but today we know that isn’t yet possible."

Claude analyzed usage data through the Datadog MCP server to identify the four highest-impact user journeys. It then estimated the impact of about twenty hand-picked projects in milliseconds, allowing the team to set realistic sprint targets. By day three, the team had hit twelve of the thirteen targets. Key technical improvements included baking a static composer into the HTML to allow typing during React initialization, precompiling a V8 code cache for the desktop shell, and cutting sidebar re-renders by 90%.

When initial targets were exceeded, the team prompted Claude to identify new opportunities, asking for "WACKY ideas" and exploring unexplored areas. This iterative process allowed the team to iterate faster than their deploy cadence, leveraging Claude’s ability to work asynchronously and validate prototypes without waiting for field data.

What It Means for Developers

This case study offers a concrete example of how AI agents can be integrated into performance engineering workflows. The key takeaway is the shift from measurement as a preliminary step to measurement as an active part of the optimization loop. By providing deterministic, countable metrics, developers can give AI agents clear, verifiable goals. This approach reduces the risk of the agent optimizing for the wrong metric, a common pitfall in automated performance tuning.

For developers looking to implement similar workflows, the Anthropic team’s use of Valgrind and V8 coverage for deterministic counting is a practical technique. It transforms noisy latency data into stable, comparable numbers that can be used in CI pipelines. This method is particularly useful for JavaScript-heavy applications where traditional profiling tools may be too slow or noisy for continuous integration.

What to Watch

• Adoption of Deterministic Benchmarks: Look for other engineering teams adopting similar deterministic counting methods for AI-driven optimization.
• Autonomy Levels: Monitor how Anthropic and other companies expand the autonomy of AI agents in production environments, particularly in high-stakes areas like performance and reliability.
• Tooling Integration: Watch for the development of MCP servers and other tools that facilitate AI interaction with observability platforms like Datadog and CI systems.
• Incident-Free Scaling: The claim of 3,000 changes with zero incidents is a strong data point; watch for similar results or counterexamples in other AI-assisted development workflows.