What was compared
The review handed the same task to GPT-5.6's Work and Claude Fable 5's Cowork: build a dashboard that pulls a YouTube channel's data live and refreshes on reload. Both tools recently opened a work-collaboration space (Claude moved desktop Cowork to the web; GPT merged Codex into the app), and the test compared completion quality and token usage under identical conditions.
Token usage was similar (Claude around 5%, GPT around 8%). So the outcome hinged not on consumption but on approach and the character of the result.
How the two tools diverged in temperament
Claude got what you wanted, fast, by any means. To reach YouTube it had the user paste an API key straight into the chat (accepting the security risk) and finished the dashboard immediately as a live HTML artifact. The take: it suits situations where you need a result quickly, like an urgent report or meeting material.
GPT thought long and played it safe. It authenticated via OAuth, spun up three design options as subagents before starting, and treated the API key as a risky value, storing it as a secret. On top of that, its new Site feature went beyond the frontend to produce an actually deployable service (domain connection, environment variables included) — as if a deploy tool like Lovable had moved inside the chatbot.
The real shift — the harness gets absorbed
The most striking part isn't the output but the structural change. GPT's Work has subagents built in, so the user never even needs to know the concept exists. The moment it spun up design options in parallel at the start is one example.
We used to wire skills and subagents into a harness by hand. That part has melted into the tool. It's a scene where the observation that 'the harness becomes unnecessary, the LLM absorbs the harness itself' is starting to show up in a shipping product.
Why it matters for AX
The biggest wall in corporate AX training isn't workflow design — it's whether people can actually operate the tool. No matter how much you say the workflow must change, nothing happens if people can't handle the tool. And a CLI like Claude Code or Codex intimidates practitioners, because you have to teach subagents and skills alongside it.
A Work-style surface with the harness absorbed inside, by contrast, is something you can recommend and teach practitioners right away. That's the context in which the reviewer gave GPT the slight edge — for teachability. The conclusion isn't a winner but a division of labor: fast Claude for the urgent report, careful GPT for the big deliberate project. We're moving into a stage of splitting the work by each tool's temperament.