이미지: @ClaudeDevs (X) 영상 갈무리
Summary
- Anthropic's developer account says it rebuilt the streaming renderer for Claude's web and desktop apps, making long answers roughly 4x smoother to read
- On low-end laptops, freezing occurs 9x less often and the worst freezes are 4.5x shorter, while 120Hz MacBooks maintain 120fps all the way to the end of a response
- The change comes from redrawing only the parts of the screen that are still changing, and it applies automatically to both the web and desktop apps with no settings to adjust
- 발표
- ClaudeDevs(앤스로픽 개발자 계정), 2026년 8월 24일
- 핵심 변경
- 스트리밍 렌더러를 재구축해 변경 중인 부분만 다시 그림
- 체감 개선
- 긴 답변 스트리밍이 약 4배 매끄러워짐
- 저사양 노트북
- 정지(stall) 현상 9배 감소
- 최악 프리즈
- 지속 시간 4.5배 단축
- 120Hz 맥북
- 답변 시작부터 끝까지 120fps 유지
- 적용 범위
- 클로드 웹, 데스크톱 앱
Fewer moments where scrolling used to stutter
When you ask Claude a long question, the answer streams onto the screen character by character. The longer the response, the more likely you've noticed that flow break up along the way. This was especially true on lower-spec laptops, where the screen would briefly freeze as more text piled up. According to a post from Anthropic's developer news account, ClaudeDevs, the team fixed this, and long answers now stream roughly 4x more smoothly on both the web and desktop apps.
What changed
The original streaming renderer appears to have worked by redrawing the entire screen every time a new chunk of text arrived. That was fine for short answers, but as responses got longer, the area that needed redrawing kept growing, and the browser or app struggled to keep up. This time, the team rebuilt the renderer so it only redraws the parts of the screen that are actually changing. Everything else stays put, which means the amount of work needed per redraw doesn't keep scaling up no matter how long the answer gets.
The numbers behind the improvement
The published figures break down into three parts, covering both stability on low-end hardware and frame rates on high-end hardware.
| Metric | Improvement |
|---|---|
| Overall streaming smoothness | ~4x |
| Freeze frequency on low-end laptops | 9x reduction |
| Worst-case freeze duration | 4.5x shorter |
| Frame rate on 120Hz MacBooks | Holds 120fps from start to finish |
The 9x and 4.5x figures were both measured on the same low-end laptop, but they track different things: how often freezing happens versus how long each freeze lasts. 120fps means the screen redraws 120 times per second — the kind of smoothness usually associated with gaming monitors, now sustained throughout a text response as well.
How to check it out now
This isn't a toggle you need to turn on. It's a change to the rendering approach itself, so simply opening Claude's web or desktop app as usual and asking a long question should let you feel the difference right away. It's most noticeable with requests that fill the screen multiple times, like long explanations packed with several code blocks or draft documents. That said, this announcement covers the web and desktop apps specifically; there's no mention of the mobile app.
One user responded to the news with a joking comment: "Good but will it also start 'making sense' while answering?" — a nod to the fact that response speed is a separate issue from how logically coherent the content actually is.
Editor's take
What's interesting about this update is that it leaves the model itself untouched and only changes how the screen renders. When you use a model like Claude Opus 5 for real work, where answers tend to run long, people's complaints often have less to do with how "smart" the model is and more to do with how comfortable the screen feels while they wait. For anyone streaming in a long code review or document draft, every little freeze chips away at trust in the tool. Rendering optimizations don't get announced with the fanfare of benchmark scores, but for people using a product every day, they often matter more than a model upgrade does.
For teams in Korea using Claude for writing documents or generating code, the practical point here is that this improvement arrives with no extra work required. There's no new API to learn or setting to change — it just layers on top of the existing workflow. Teams working on lower-spec office laptops are likely to notice the 9x drop in freeze frequency the most.
In the coming weeks, it wouldn't be surprising to see OpenAI or Google announce similar rendering optimizations. The chatbot competition already seems to be shifting toward a new kind of felt metric beyond raw model performance: how smoothly, and without interruption, an answer actually appears on screen.
Comments