Best AI agent scores just 12% on real professional work

The latest:

The strongest AI agent tested solved only 12.0% of realistic professional tasks, according to GDPx, a benchmark SemiAnalysis launched on 2 November 2026 covering 11 domains from chip design to quantitative trading. Version 1.0 contains 206 tasks built with 14 specialist data partners and ranks 13 models. Anthropic’s Claude Opus 5.5 placed first, ahead of Gemini 4 Carbon and GPT-6 Astra.

Details:

• The leaderboard: Claude Opus 5.5 ranked first at 12.0%, averaging 51.41 dollars and 548,000 output tokens per task. Gemini 4 Carbon came second at 10.9% for 16.29 dollars per task, GPT-6 Astra third at 9.6% for 53.99 dollars, Meta Muse Wonder Full fourth at 9.0%, and Z.ai GLM 5.3 last of 13 at 2.7%.
• The domains: GDPx spans chip design, cybersecurity, biology, finance, law, quantitative trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis said it excluded government, real estate and insurance, arguing the first trillion dollars of AI revenue will concentrate in a few sectors.
• The grading: Scores are mean pass@1 over 3 attempts, with each domain weighted equally and each task carrying its own binary or graded rubric. Finance and law returned zero for every model under binary grading, so SemiAnalysis publishes partial-credit averages for those two domains instead.
• Task realism: Every task runs in a live environment with a default 8-hour budget. Finance tasks begin from a data room of roughly 4,000 files drawn from a real, anonymized private equity deal — a scale closer to a junior analyst’s workload than to a single prompt.
• Safety effects: SemiAnalysis said safety classifiers depressed scores for Opus 5.5 and GPT-6 Astra in cybersecurity, biology and alignment. On an adjusted 8-domain ranking excluding those areas, GPT-6 Astra edges Opus 5.5 by 0.1 points and Gemini 4 Carbon falls to fifth — an order the firm said is closer to its internal assessment.
• Harness choice: Each model ran inside its own lab’s agent tooling: Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon, with OpenCode as a fallback. Named data partners include Aria, Arcverex and Ergodex on chip design, Living Context on finance and LatchBio on biology.
• The rival benchmark: SemiAnalysis positions GDPx as a successor to OpenAI’s GDPval, which it criticizes because the tasks are non-interactive and expert agreement reaches only 71%. Its partners cleared an automated quality check on the first attempt, and the firm said that same check flags roughly 50% of Terminal Bench 4.0 tasks as broken.
• Trading results: In quantitative trading, 70 of 111 graded runs came in below the reference strategy, and no model solved the general aapl-target-winsorization task — one of the clearest single data points on where current agents break down.

Background:

GDPval, released by OpenAI, measured model output against economically valuable work but used static, non-interactive tasks. GDPx keeps the economic framing while running agents inside working environments with tools, files and multi-hour budgets.

Between the lines:

The 12.0% ceiling and the zero scores in finance and law under binary grading point the same way: agents produce partial work on long professional tasks rather than finished deliverables. Cost spreads reinforce it — Gemini 4 Carbon scored within 1.1 points of the leader at under a third of the per-task cost, so price, not capability alone, may decide deployment.

What’s next

SemiAnalysis describes GDPx as a living benchmark with new versions every 3 to 6 months, making the next release the first test of whether scores move. Watch whether finance and law stay at zero under binary grading.