How close is AI to experts?

By sector, September 2025

The paper's appendix breaks the seven models out by sector. No later launch post does, so this is the only sector view there is. Claude Opus 4.1 was ahead in eight of the nine (by a point in Wholesale Trade); GPT-5 led in Information. In three sectors the best model reached 50%.

What parity here does not mean

Matching an expert on a GDPval task is not the same as doing the expert's job. The paper and OpenAI's own write-up are plain about why:

• The tasks come fully specified, and in one shot. The prompt hands the model the whole context and the reference files. Real work starts with working out what the task is. When the paper cut the prompts to 42% of their length, GPT-5 did worse, because it struggled to find the context it needed.
• Only computer work, and not all of it. The paper leaves out physical tasks and anything needing tacit knowledge, personal data, proprietary software, or talking to other people. The 44 occupations were picked because at least 60% of their O*NET tasks are digital.
• A job is more than its tasks. OpenAI's announcement says it directly: "most jobs are more than just a collection of tasks that can be written down."
• Someone still has to check the work. Raw model time was about 90 times faster than the expert for GPT-5. Once the paper added an expert's review time (about 109 minutes a task) and redoing the work when the model lost, the gain fell to 1.12 times, or 1.39 times if the expert resampled the model until it worked.
• Losses vary in kind. When GPT-5 lost to the expert, other graders called the model's work acceptable but weaker most often; about 29% of ratings were "bad" or "catastrophic", about 3% "catastrophic", which the paper defines with examples like a wrong diagnosis or an insulted customer.
• Longer tasks go worse. Win rates were highest on tasks that take an expert 0 to 2 hours and fell as tasks got longer. The public tasks average 9.5 hours of an expert's time, by the experts' own estimates.
• Graders disagree with each other. Two experts grading the same pair agreed 71% of the time. And the comparisons were not perfectly blind: the paper notes OpenAI outputs used em dashes, Claude wrote in the first person, and Grok sometimes named itself.
• Small samples. Each occupation has five public tasks. The per-occupation bars rest on 45 graded comparisons per model, and the paper's error bars on them are wide.

The chart has its own caveats. Every expert-graded score after the paper comes from OpenAI announcing its own models, run in its own setups (more reasoning effort, new ChatGPT tools, "custom skills" for GPT-5.3-Codex), and no one else has published an expert-graded run. I checked the launch posts and model cards of Anthropic, Google, xAI, Meta, DeepSeek, Moonshot, Z.ai, Alibaba and MiniMax since September 2025: they quote Artificial Analysis' Elo (below), or in MiniMax's case its own LLM-judged variant, never a rate against the experts. Claude Opus 4.7 and Gemini 3.1 Pro appear only because OpenAI printed them in a comparison table without saying who ran them. After GPT-5.5, OpenAI stopped reporting this number and switched to Artificial Analysis' Elo, and OpenAI's own GDPval leaderboard at evals.openai.com now says it is no longer active and points there too.