Jessica Wachtel at The New Stack ran three coding-work tests through Claude Cowork and ChatGPT Work mode and reported ChatGPT 22 percent lighter on tokens while Claude added detail on every task

Anthropic merged Claude chat and Cowork on 22 September 2026, and OpenAI has offered the same shape since summer as ChatGPT Work mode. Jessica Wachtel at The New Stack put both through three real developer tasks the same day and reported every accuracy number tied. The differences she found are on speed, output tokens and the amount of context each agent added on top of the correct answer.

The three tests and their numbers

Wachtel ran an API rate-limit lookup, a code-from-spec task, and a stack-trace plus log analysis, and recorded time and token cost on each. On the rate-limit table, "both Claude and ChatGPT answered correctly on the twelve graded fields, but Claude was more thorough". ChatGPT finished in 1 minute 17 seconds and 649 output tokens. Claude took 1 minute 44 seconds and 1,042 tokens, "read eight pages, listed nine sources, and offered to export the table as a spreadsheet".

On the code task, Wachtel reports the code came out "nearly identical, both using exact-precision arithmetic and a fixed-order regex". Claude flagged that rounding 0.5 seconds up is a judgment call the spec never settles, and Python's built-in round would go the other way. ChatGPT reported only that its tests passed.

The third test started as a stack-trace question and handed off to a log-analysis task in the same thread. On the trace, both answered correctly in seconds. On the log build, Wachtel writes that "Claude pulled the file, computed the table, and built an .xlsx with a p95 bar chart in about 4 minutes on 319 output tokens", while "ChatGPT produced the same spreadsheet and chart in 28 seconds, using 467 output tokens", with "all 30 numbers matched my ground truth on both sides". Claude named its percentile method as linear interpolation; ChatGPT did not.

The overall split

Across the three tasks Wachtel reports 1,933 output tokens for ChatGPT and 2,468 for Claude, or 22 percent lighter for ChatGPT. ChatGPT was faster on every task. On the short tasks the gap was seconds (1 minute 17 seconds against 1 minute 44 for one, 1 minute 17 against 1 minute 45 for the other). Wachtel writes that "some of Claude's extra output was filler, but not all of it. It added context the prompts didn't ask for, named the percentile method behind its numbers, and flagged two judgment calls the specs left open".

Her pick was Claude, on the argument that when accuracy ties, the answer that names its assumptions is the one that survives an unfamiliar codebase. That is a judgment call in a review by one person on three tasks, not a benchmark, and Wachtel included her prompts so anyone with an account on both services can replicate it.

What this changes for a team

For a team choosing between the two agents, the useful takeaway from Wachtel's tests is that the two sides now handle real developer work at comparable accuracy, and the split is on the shape of the answer rather than on which one gets it right. A team that grades an agent on speed and token cost has different numbers to look at than a team that grades on whether an agent surfaces the judgment call a spec left open. Wachtel's numbers are worth taking as a starting point for a team's own eval on the tasks they actually run, rather than as a purchasing decision on their own.

Source

Jessica Wachtel, The New Stack, Claude's merged chat and Cowork vs. ChatGPT's Work mode: ChatGPT is faster, Claude is more thorough, 22 September 2026.