1. Home
2. News
3. Anthropic

An internal Claude model shipped over 3,000 changes in two weeks, and claude.ai got 3x faster

Anthropic's engineers approved every change, and a lot of their job was telling the model to be bolder.

By Think Facility · September 24, 2026 · Anthropic AI agents

On September 23, 2026, three Anthropic engineers wrote up how they made claude.ai and the Claude desktop app about 3x faster in a two-week sprint this August. Claude did most of it. They ran an internal research model (roughly comparable to Opus 5.5, they say) out of one Slack channel.

The model found the slow spots and built the benchmarks. It wrote the fixes, then watched each deploy. The engineers set goals and approved every change, and they merged more than three thousand of them without a customer-facing incident or a rollback.

Sprint

Two weeks in August

Changes merged

More than three thousand

Incidents

Without a single customer-facing incident or rollback

Model

An internal research model roughly comparable to Opus 5.5

What got faster

They picked four journeys that make up 95% of what people do in the apps. Opening the app is one; sending a message is another. At the 75th percentile, a fresh load of claude.ai used to take 3.1 seconds before you could type. Now it's 0.55. A new Claude Code session starts in 0.3 seconds, down from 0.8.

They'd set thirteen targets. Twelve were hit by day three, so they set new ones. At the peak, more than a hundred and fifty threads ran at once, each chasing a single benchmark. On the busiest days over two hundred changes landed.

My favorite find is a tiny one. Highlighting a finished code block could freeze the page for about a second. Claude traced it to em dashes. Turns out any character outside Latin-1 made the browser's JavaScript engine store the whole reply in a slower format, and a twenty-line change fixed it.

The model had to be pushed

What surprised me is how much time the engineers spent asking the model to take bigger swings. Left alone, it kept shrinking its plans and padding its estimates.

In one thread Claude offered to put a change up "this week". Raymond Wang wrote back that he'd get it merged and deployed if it went up right now. Be braver, he said. The PR was promised within the hour.

Granted, nobody's claiming it ran itself. Every thread had a named human owner and every PR needed at least one human approval. Anything a user could see shipped behind a feature flag. In their words, "The loop was productive, but it wasn't autonomous."

Boris Cherny (the post thanks him for pushing the team to be more ambitious) pointed other engineers at the write-up that afternoon.

What's next

The team says the 95th percentile, other journeys and very long conversations still have room to improve. It's planning a separate post on fixes that went upstream into Electron, Chromium and Node.js during the sprint.