SemiAnalysis released GDPx on November 2, 2026 — a new benchmark that gives frontier AI agents realistic expert jobs across 11 professional domains — and the best model on its first leaderboard, Anthropic’s Claude Opus 5.5, completed just 12.0% of the work. Why it matters: GDPx is built to measure the kind of long, messy, multi-step work companies actually want to automate, and on that measure every frontier model still fails most of the time.
November 2, 2026
What does GDPx test?
GDPx v1.0 contains 206 tasks built with 14 specialized data partners, spanning chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis says it picked domains where it believes the first trillion dollars of AI revenue will concentrate, and deliberately left out areas such as government, real estate and insurance. Tasks carry a default time limit of eight hours, and models run in their native harness where one exists — Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon — with OpenCode as the fallback. The headline score averages each model’s pass@1 result across three attempts, with every domain weighted equally.
How did the models score?
Claude Opus 5.5 leads at 12.0%, at an average cost of $51.41 per task. Google’s Gemini 4 Carbon is second at 10.9% and far cheaper per task at $16.29, OpenAI’s GPT-6 Astra is third at 9.6% at $53.99 per task, and Meta’s Muse Wonder Full is fourth at 9.0%. No model dominates: Opus 5.5 won quant trading, hardware design and chip design; Gemini 4 Carbon led biology and narrowly won software engineering; GPT-6 Astra won recursive self-improvement and materials science. In finance and legal, every model scored zero under the benchmark’s intended pass-or-fail grading, so SemiAnalysis reports partial scores for those two domains instead.
What are the caveats?
Two, and SemiAnalysis states both itself. First, it says the Opus 5.5 and GPT-6 Astra overall scores are pulled down by their safety classifiers in the cybersecurity, biology and alignment domains, and it describes Muse Wonder Full as the best available cyber model by a meaningful margin. Second, these are SemiAnalysis’s own results — no outside group has re-run them yet. The benchmark is also a deliberate shot at OpenAI’s GDPval, released in September 2025, which the GDPx authors criticize for handing models all the context they need inside the prompt; GDPx tasks instead ship with a realistic working environment.
Why it matters
For anyone buying AI agents on the strength of launch-day benchmark charts, GDPx is a useful cold shower. The tasks here look like real expert work — long time limits, real file environments, pass-or-fail grading — and the best system available finishes about one job in eight, at roughly $51 a task for the leader. SemiAnalysis plans to update GDPx every three to six months as a living benchmark. If that cadence holds, its leaderboard will be one of the more honest progress meters in AI: everything AI, tested, against work that is actually hard.
Frequently asked questions
Which AI model scored highest on GDPx?
Anthropic’s Claude Opus 5.5, at 12.0% overall, followed by Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%.
Why are the GDPx scores so low?
The tasks are realistic expert jobs with up to eight hours of work, graded pass-or-fail in most domains. In finance and legal, all 13 ranked models scored zero under that binary grading.
Have the GDPx results been independently verified?
No. The scores are SemiAnalysis’s own; as of publication, no outside group has re-run the benchmark.
Sources: SemiAnalysis (GDPx announcement and leaderboard), Unite.AI, Crypto Briefing, AI Weekly