Claude’s merged chat and Cowork vs. ChatGPT’s Work mode: ChatGPT is faster, Claude is more thorough

When Anthropic merged Claude chat and Cowork into a single interface last week, users no longer had to choose between the two modes manually. Now, users can simply ask a question and hand off a multi-step task within the same thread, and Claude will handle the rest. Anthropic stated that customers often had difficulty selecting the appropriate tab for each task, so the merged app now automatically routes each request.

OpenAI had introduced a similar concept with its Work mode, which was launched alongside Chat on July 9 and later replaced the separate Agent mode the following month. Work mode provides a sandboxed environment with a browser, code execution, and file output alongside a Chat toggle in the same window. However, it cannot currently seamlessly return a live, logged-in browser session mid-task, similar to Agent mode.

OpenAI built Work mode on its Codex coding agent, noting that about one-fifth of Codex's 5 million weekly users were non-developers, a share that was growing three times faster than the developer segment.

In a recent head-to-head test, the two AI models were evaluated based on their accuracy, reliability, and token usage in various developer-related tasks. Claude's ability to provide more thorough responses was evident in all three tests.

1. API Research: Claude was able to tabulate rate limits, free tier information, and API version identifiers for four real developer APIs, including GitHub REST API, Stripe API, Twilio Messaging API, and OpenAI API. It included additional details such as GitHub's separate limit for Actions tokens, Twilio's queue window, and OpenAI's tier thresholds, which ChatGPT did not cover.

Claude took longer to complete the task (1 minute 44 seconds) and wrote more tokens (1,042) compared to ChatGPT (1 minute 17 seconds, 649 tokens), but the added detail was warranted due to the increased complexity of the task.

2. Build from a Spec: Claude was asked to write a command-line duration parser based on a strict specification. It completed the task in 1 minute 45 seconds and wrote 989 output tokens, while ChatGPT finished in 1 minute 17 seconds and produced 769 output tokens. Although both delivered a working Python file (durparse.py), Claude included an additional test case and demonstrated a slightly different approach in handling certain edge cases, such as rounding 0.5 seconds up, which ChatGPT did not explicitly mention.

3. Handoff: In the final test, Claude was asked to identify which stack trace indicated a race condition and then analyze a CSV file containing log data, compute latency percentiles and error rates per endpoint, and deliver a spreadsheet with a bar chart. Claude completed this task in 4 minutes and 319 output tokens, while ChatGPT managed the same task in 28 seconds using 467 output tokens.

Claude also provided additional information about its percentile calculation method and noted that the numbers would align with those obtained by recomputing them using Google Sheets.