SemiAnalysis published GDPx on November 2, 2026, describing it as a general agentic benchmark aimed at domains where frontier models are most economically valuable. Authors Max Kan, Andrew Megalaa, and Dylan Patel say even the best model scores just 12%, which they treat as evidence the suite is still frontier-hard rather than saturated.
The official site lists 11 domains, 14 partners, and 206 tasks. Claude Opus 5.5 tops the v1.0 leaderboard at 12.0%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.
Key Facts
| Item | Detail |
|---|---|
| Release | November 2, 2026 (SemiAnalysis newsletter) |
| Scale | 11 domains, 14 partners, 206 tasks |
| Headline score | pass@1 averaged over three attempts; domains weighted equally |
| Top 3 | Claude Opus 5.5 12.0%; Gemini 4 Carbon 10.9%; GPT-6 Astra 9.6% |
| Also listed | Muse Wonder Full 9.0%; Claude Fable 5.1 8.5%; GPT 6.1 Sol 8.4%; down to GLM 5.3 2.7% |
| Default time limit | 8 hours per task |
| Harness | Native when available (Claude Code, Codex, Antigravity); OpenCode otherwise |
- Domains. Chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering. SemiAnalysis intentionally skips government, real estate, and insurance-style GDP share taxonomies.
- Partners (named in the post). Aria, Arcverex, Ergodex (chip); Arithmetic (cyber); Latch Bio (bio); Living Context (finance); Chakra (legal); Third Moment (quant); Extrapolate (RSI); Beautiful Machines (hardware); ValenceForge and Emulated (materials); Fleet (alignment); Datacurve (software).
- Domain winners (authors). Opus 5.5 wins quant trading, hardware design, and chip design. Gemini 4 Carbon leads biology and edges software engineering. GPT-6 Astra leads RSI and materials science. Finance and legal return 0 under intended binary grading for every model tested; SemiAnalysis reports average partial scores for those domains instead.
- Safety classifiers. The authors say Opus 5.5 and GPT-6 Astra overall scores are sandbagged in cyber, biology, and alignment because safety classifiers fire. They call Meta's Muse Wonder Full the best available cyber model by a meaningful margin in their results.
- Cost notes (Unite.AI / site). Unite.AI cites Opus at about $51.41 average cost per task, Gemini 4 Carbon about $16.29, and GPT-6 Astra about $53.99. The site's Pareto chart omits Muse Wonder Full for lack of a public price.
Background
SemiAnalysis frames GDPx as a successor critique of OpenAI's GDPval (September 25, 2025). The authors argue GDPval tasks are non-interactive and spoon-feed context in the prompt, with human expert inter-rater agreement of 71% on its gold subset. GDPx tasks instead ship with realistic environments. One finance example reconstructs a private-equity deal (about 62% of a snack company for $880 million plus earnout) across roughly 4,000 anonymized files.
Partners had to pass SemiAnalysis automated QA on the first try with only minor mistakes. The same process, they write, flags nearly 50% of Terminal Bench 4.0 tasks as broken. GDPx is planned as a living suite with new versions every 3–6 months.
Reactions and Impact
Unite.AI's November 2 report repeats the 12.0% / 10.9% / 9.6% leaderboard and partner list. The GDPx site exposes domain comparisons and rollout pages for one public task per partner. As of publication, major lab blogs had not issued formal rebuttals in the sources we checked; the primary artifact remains the SemiAnalysis post and leaderboard.
What We Don't Know Yet
- How labs will treat GDPx on future model cards versus GDPval
- Full task counts and private-task distributions per domain beyond the public examples
- Whether safety-classifier-adjusted rankings will become an official alternate board
- Independent third-party replications of the reported scores