Claude's merged app now routes each request between chat and Cowork itself

Work mode is built on OpenAI's Codex agent ^[9]. OpenAI said roughly a fifth of Codex's 5 million weekly users were non-developers, and that the non-developer share was growing three times faster than the developer one ^[10]. A fifth of 5 million is about a million people running a coding agent without writing code ^[4]. What they get is a sandbox with a browser, code execution and file output, and a Chat toggle beside it in the same window ^[5]. The toggle is still a switch the user throws.

Anthropic took the switch out. The company says customers often struggled to choose the right tab for the right task, so the merged app routes each request itself ^[3]. Ask a question, hand off a multi-step task in the same thread, and Claude decides what runs it ^[2]. The New Stack's account does not report whether a user can pin a request to one mode.

The numbers come from one reviewer's three tasks, run once each, with the prompts published ^[11]. On the API research test, both apps answered all twelve graded fields correctly, and Claude added GitHub's separate limit for Actions tokens, Twilio's queue window, OpenAI's tier thresholds and a note about a page it could not reach ^[12]. ChatGPT, in Work mode, finished in 1 minute 17 seconds on 649 output tokens ^[13]. Claude took 1 minute 44 seconds and 1,042 tokens, read eight pages and listed nine sources ^[14].

The New Stack writes that Claude "wrote nearly double the tokens and took longer" ^[15]. 1,042 against 649 is about 61 percent more ^[1]. The time gap is 27 seconds, about 35 percent ^[2]. On the build-from-a-spec test the gap narrows to 989 tokens against 769, about 29 percent, with 28 seconds more on the clock ^[18][19][3]. Part of Claude's extra spend is accounted for: it reported running 35 of its own test cases before returning the file, against the 16 hidden cases the reviewer then ran ^[20][5]. The two files came out nearly identical, both using exact-precision arithmetic ^[21].

For either set of timings to transfer, your prompts would have to sit near the 517-token spec in size ^[17] and your grading would have to reward the same added detail. One run per task also leaves the variance unmeasured.

The third test is the one aimed at the merged thread: identify which of two stack traces shows a race condition, then in the same conversation hand off a job that pulls a log file from Google Drive, computes latency percentiles and error rates per endpoint, and returns a spreadsheet with a chart, against log data the reviewer generated so every number was known in advance ^[22]. ChatGPT posted the same 1 minute 17 seconds on both scored tests ^[6]. I would want that repeated ten times before reading it as fixed per-run overhead.