By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
4 min read
Key Takeaways
SemiAnalysis published GDPx on November 2, a benchmark that grades AI agents on 206 expert tasks across 11 domains. On its leaderboard, Anthropic's Claude Opus 5.5 scored highest at 12.0%; Google's Gemini 4 Carbon was second at 10.9% and OpenAI's GPT-6 Astra third at 9.6%. SemiAnalysis says its finance and legal tasks are so challenging that “basically every model scores exactly 0.”
12%
That is the top score on GDPx, a new benchmark for general agentic capability that the research firm SemiAnalysis published on November 2, 2026. The authors — Max Kan, Andrew Megalaa, and Dylan Patel — write: “GDPx is a true frontier-level benchmark: even the best model scores just 12%.”
The debut leaderboard ranks 13 models. Anthropic's Claude Opus 5.5 leads at 12.0%, at an average cost of $51.41 per task. Google's Gemini 4 Carbon — not yet generally available — is second at 10.9% and $16.29 per task. OpenAI's GPT-6 Astra is third at 9.6% and $53.99 per task. Meta's Muse Wonder Full is fourth at 9.0%. The best open-weight model on the board, Z.ai's GLM 5.3, scores 2.7%.
SemiAnalysis says the headline score is pass@1 averaged over three attempts, with each of the 11 domains weighted equally. Tasks carry a default eight-hour limit, though some are shorter, and each model runs in its own native harness where one exists — Claude Code for Opus, Codex for Astra, Antigravity for Carbon — with OpenCode as the fallback.
How GDPx is built
SemiAnalysis contrasts GDPx with OpenAI's GDPval, released September 25, 2025, which it calls the most popular general-purpose benchmark still reported on major model release cards. SemiAnalysis writes: “Every GDPval task is precisely specified and non-interactive, with all the necessary context directly spoon-fed to the model in the prompt.”
The 206 tasks span 11 domains that SemiAnalysis describes as the most valuable for frontier intelligence: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. SemiAnalysis says it intentionally left out industries such as government, real estate, and insurance. Fourteen specialist data partners supplied the tasks, according to SemiAnalysis — Latch Bio for biology, Datacurve for software engineering, Chakra for legal, Third Moment for quant trading, and others.
SemiAnalysis says the finance tasks are seeded by a real transaction: a private-equity fund, Bannock Capital, buying about 62% of a family snack manufacturer, Haverly Snack Company, for $880 million plus an earnout, spread across roughly 4,000 anonymized files. The agent has to find the right documents in the deal room before it can do the work. According to SemiAnalysis, cyber tasks are built from proprietary zero-day vulnerabilities in widely deployed open-source software, and materials tasks put the model inside simulated labs with battery cyclers and a transmission electron microscope.
SemiAnalysis says every partner had to pass its automated quality checks on the first try, with only minor mistakes at worst, and that running the same checks on Terminal-Bench 4.0 flags nearly 50% of its tasks as broken.
Results by domain
SemiAnalysis writes that “Opus 5.5 is the best all-around model and wins in quant trading, hardware design, and chip design.” It says Gemini 4 Carbon is the best biology model and edges out the win in software engineering, and GPT-6 Astra wins recursive self-improvement and is the best model for materials science.
On finance and legal, SemiAnalysis writes: “Our finance and legal tasks are so challenging that basically every model scores exactly 0.” SemiAnalysis says it also reports average partial scores.
SemiAnalysis also says Opus 5.5 and GPT-6 Astra have “sandbagged overall scores due to their safety classifiers,” which sometimes kick in on cyber, biology, and alignment tasks. It adds that Meta's Muse Wonder Full, which refuses less, is “actually the best available cyber model today by a meaningful margin.”
On the leaderboard, Opus 5.5 averages $51.41 per task, roughly three times Carbon's $16.29. GPT-6.1 Sol scores 8.4% — close to the leaders — at $9.80 per task, and DeepSeek's 4.1 Flash scores 5.4% at $0.94.
Caveats and what comes next
GDPx is SemiAnalysis's own benchmark, and the published scores are from SemiAnalysis's own runs. Gemini 4 Carbon is not yet generally available, SemiAnalysis notes. SemiAnalysis says it plans to release new versions of GDPx every three to six months.