Most AI benchmarks test exam questions: maths problems, multiple-choice science, short coding puzzles. Companies, however, pay people for deliverables: a financial model, a care plan, a sales deck, a legal memo. GDPval was built to measure exactly that. It asks AI models to produce real work from 44 occupations and compares the result with what professionals produce. In November 2026, the independent GDPval-AA leaderboard puts Anthropic's Claude Opus 5.5 at the top (Elo 1866), statistically level with China's Kimi K3.1 Pro and Meta's Muse Wonder. OpenAI's best entry is 247 points behind, Mistral's 435. This guide explains how GDPval works, what the latest numbers say about Anthropic, OpenAI, Mistral and the Chinese labs, what a task costs, and what the benchmark does not tell you.
What GDPval measures
OpenAI published GDPval in September 2025 (the paper was later accepted at ICLR 2026). The dataset has 1,320 tasks from 44 occupations in the nine sectors that each contribute more than 5% of US GDP, from finance and healthcare to manufacturing, retail and government. The tasks were written by professionals with an average of 14 years of experience and come with reference files: spreadsheets, PDFs, images, briefs. The model must return a finished work product, such as a Word document, an Excel model, a slide deck, a schematic or a stage plot. A public "gold set" of 220 tasks is what most outside evaluations use.
In the original study, experts from the same occupation compared the AI deliverable with a deliverable made by a human expert, without knowing which was which, and chose the better one. The headline number is the share of tasks where the model wins or ties. 50% means the model is as good as the expert on average.
Because expert grading is slow and expensive, Artificial Analysis built GDPval-AA, an independent version that runs continuously. Every model gets the same agent harness (Stirrup) with shell access and web browsing, works on the 220 gold tasks and produces files. An AI judge then compares the outputs of two models blind, and the results are turned into Elo ratings, anchored so that DeepSeek V4.1 Flash at maximum effort equals 1600. Version 2.1 currently rates 319 model configurations.
How to read Elo. The difference between two ratings gives the expected share of head-to-head wins. 0 points is 50/50, 100 points is about 64%, 200 points about 76% and 400 points about 91%. So Claude Opus 5.5 (1866) would be expected to produce the better deliverable in about 81% of comparisons against OpenAI's best entry (1619) and about 92% against Mistral Large 4 Preview (1431). Against Kimi K3.1 Pro (1861) it is a coin toss: the 95% confidence intervals of the top three overlap.
The leaderboard: Anthropic leads, Chinese labs are level, OpenAI trails
The chart shows the best configuration from each developer on GDPval-AA v2.1. Anthropic does not only hold first place: nine of the top 15 entries are Claude models (Opus 5.5, Sonnet 5.5, Fable 5.2, Fable 5.1 and Opus 5 at different effort levels). Close behind are Kimi K3.1 Pro from Moonshot AI (1861), which is an open-weight model, and Meta's Muse Wonder Full (1857), which has no public API. Alibaba's Qwen4 Max (1809), released this week, is fourth among developers.
Source: Artificial Analysis, GDPval-AA v2.1 leaderboard, accessed 19 November 2026. Best configuration per developer (Claude Opus 5.5, Muse Wonder Full, DeepSeek V4.1 Flash 1017 and GPT-5.6 Sol at max effort; Gemini 4 Carbon at high; Grok 4.7 at high). Confidence intervals are about ±15 to ±28 points. Amber: Anthropic, blue: OpenAI, red: Mistral, dark grey: Chinese labs.
The surprise is OpenAI. GDPval is OpenAI's own benchmark, yet its best entry on the independent leaderboard is GPT-5.6 Sol from July (1619). The newer GPT-6.1 Sol scores 1607 (42nd place) and GPT-6 Sol 1514 (62nd). OpenAI's own measurements look different (see below), which shows how much the evaluation setup matters: GDPval-AA runs every model in the same neutral harness, not in ChatGPT's own agent environment. The newest GPT-6.2 Sol has not been rated yet.
Mistral, Europe's leading lab, is far behind. Mistral Large 4 Preview (October 2026) reaches 1431, in 84th place, and the earlier Mistral Medium 3.5 only 769. It is, however, a big step: more than 650 points in six months. Google is mid-table with Gemini 4 Carbon (1733). Manus does not appear at all: it is an agent product that runs on top of foundation models, not a model of its own, so it is not rated in model leaderboards like this one.
From 48% to 85% in seven months: the human-graded results
The original human-graded GDPval shows how fast the frontier is moving. In the September 2025 paper, the best model, Claude Opus 4.1, won or tied against experts on 47.6% of tasks, just below parity. GPT-5 reached 39.0% and GPT-4o, from 2024, only 12.5%. When OpenAI launched GPT-5.5 in April 2026, it reported 84.9% for GPT-5.5, 83.0% for GPT-5.4, 80.3% for Claude Opus 4.7 and 67.3% for Gemini 3.1 Pro.
Sources: Patwardhan et al., "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks" (OpenAI, Sep 2025; ICLR 2026), gold set of 220 tasks, blind grading by occupation experts; OpenAI, "Introducing GPT-5.5" (23 Apr 2026), vendor-reported. OpenAI's original blog post gave slightly different values for some models (e.g. 40.6% for GPT-5 high).
Two caveats. First, the April 2026 figures are reported by OpenAI for its own launch, with no independent replication. Second, beating an expert on a self-contained task is not the same as doing the job. GDPval's authors stress that the tasks are precisely specified and one-off; real work also means finding out what the client actually wants, iterating with colleagues and taking responsibility for the result.
How fast each lab improved
Combining release dates with GDPval-AA ratings shows when each lab reached which level. Anthropic jumped from 1240 (Claude Sonnet 4.6, February) to 1618 (Claude Fable 5, June) and 1866 (Claude Opus 5.5, September). Chinese labs closed the gap in steps: GLM-5.2 in June, MiMo-V2.6-Pro in September and Kimi K3.1 Pro in October, which pulled level with the leader. OpenAI has not improved on its July result since. Mistral made the biggest jump of the year, but from a low base.
Source: AI Workshop analysis of the Artificial Analysis GDPval-AA v2.1 leaderboard (19 Nov 2026): highest rating among each group's models released up to the given month, using today's ratings. "Best Chinese model" covers Kimi, Alibaba, DeepSeek, Z AI, Xiaomi, MiniMax, StepFun and others. Mistral's December 2025 value is Devstral 2. Only models that Artificial Analysis has evaluated are included.
Effort settings matter as much as the model
Modern models let you choose how long they "think" before answering. On professional tasks, this setting changes the result enormously. Claude Opus 5.5 scores 1239 at low effort and 1866 at maximum effort, a difference of 627 points, larger than the whole gap between Claude Opus 5.5 and Mistral Large 4 Preview (435). GPT-6.1 Sol gains 285 points from low to max. In practice, a top model at a low setting can deliver worse work than a mid-range model at a high setting. Our guide to reasoning models and test-time compute explains why.
Source: Artificial Analysis, GDPval-AA v2.1 leaderboard, 19 Nov 2026. Claude models use the "Default Fallback" configuration. Effort names follow each vendor's API (low, medium, high, xhigh, max).
What a task costs: quality per dollar
Higher effort means more tokens and more turns. Artificial Analysis also reports the average API cost of solving one GDPval task. Claude Opus 5.5 costs $8.92 per task at its top setting and averages 86 agent turns. Kimi K3.1 Pro delivers a statistically identical rating for $3.02. Claude Fable 5.2, Anthropic's largest model, is the most expensive at $26.43 and still scores slightly below Opus 5.5. At the cheap end, DeepSeek V4.1 Flash (1746) costs $0.35 and Xiaomi's MiMo-V2.6-Pro (1692) $0.16, while Claude Haiku 5.5 (1629) costs around $0.22.
Source: Artificial Analysis, GDPval-AA v2.1, cost per task at list API prices (input, cache, reasoning and output tokens), 19 Nov 2026. Models in Artificial Analysis' default selection with a public API; the Claude Haiku 5.5 cost is marked as an estimate by Artificial Analysis. Log scale on the horizontal axis.
Even the most expensive figure is small compared with the human alternative. A task that takes an experienced professional several hours costs a few hundred euros in salary; the model costs a few dollars. But the API price is not the full cost: someone has to check the result, fix errors and take responsibility. The GDPval authors make the same point: their speed and cost comparison covers only model time and API prices, not review. Our article on model routing and cost optimisation shows how to combine an expensive and a cheap model in one workflow.
What GDPval does not tell you
- An AI judge is not a client. GDPval-AA uses a model to decide which deliverable is better. That scales, but judges have biases, for example towards longer or more polished documents. See our guide to LLM-as-a-judge.
- Relative, not absolute. Elo tells you which model is better than which, not whether the work is good enough to send to a customer.
- US occupations, English tasks. The tasks reflect the US economy and are written in English. Croatian or German documents, local law and local accounting rules are not tested.
- A neutral harness. Every model works in the same agent environment. Vendors' own products, such as Claude with its skills and connectors or ChatGPT's agent mode, can perform differently, which partly explains the gap between OpenAI's own and independent results.
- Self-contained tasks. No meetings, no unclear briefs, no company context. That is exactly the part that makes real projects hard.
What this means for companies in Croatia and DACH
- Use GDPval to shortlist, not to decide. It is the best public signal for document-heavy knowledge work, but test the two or three leading models on 20 to 30 of your own tasks before you standardise. Our benchmark comparison lists other useful tests.
- Claude is the safe default for high-value deliverables. Anthropic holds the top spot and nine of the top 15 places. Which Claude model fits which job is in our Claude model selection guide.
- Set the effort level deliberately. For contracts, reports and financial models, use high or maximum effort; for routine drafts, a lower setting or a smaller model is enough.
- Take Chinese open-weight models seriously, but check data rules. Kimi K3.1 Pro and DeepSeek deliver top results at a fraction of the price. Self-hosted in the EU they can be an option; through Chinese APIs, GDPR and confidentiality usually rule them out. See our analysis of Chinese AI models.
- Mistral for sovereignty, not for peak quality. Mistral Large 4 is improving fast and can run in Europe, but on professional tasks it is clearly behind. Our Mistral guide covers where it fits.
- Plan for review, not replacement. The realistic model is AI producing the first 80% and a professional finishing and signing off. Our article on AI and the labour market looks at what that means for roles.
The bottom line
GDPval is the closest thing the industry has to a test of real professional work, and it shows a clear picture for late 2026. The best models now match or beat experts on most well-defined tasks. Anthropic leads the independent ranking, Chinese labs have caught up at much lower prices, OpenAI's independent results lag behind its own claims, and Mistral is improving from far behind. For companies, the lesson is practical: choose a top model, set it to think long enough on important work, test it on your own tasks and keep a qualified person responsible for what goes out.
Which model is best for your work?
AI Workshop tests Claude, GPT, Mistral and open-weight models on your own documents and processes, sets up the right effort levels and review steps, and builds the workflow around it. Our Claude Certified Architects start with one real deliverable and a measurable result.
Talk to an AI consultant