SemiAnalysis launches GDPx benchmark; best model scores 12% on agentic tasks

SemiAnalysis released GDPx on November 2, 2026, a benchmark of 206 agentic tasks across 11 economically valuable domains, including chip design, software engineering, finance and biology. Its headline result is that the best model scores only 12%. In a post on X, SemiAnalysis wrote that GDPx is a true frontier-level benchmark and that even the best model scores just 12%.

SemiAnalysis said it did not build GDPx around share of GDP or the SOC/O*NET occupational taxonomy. Its reasoning is that the first trillion dollars of AI revenue will concentrate in sectors that can deploy frontier intelligence at high return on investment, rather than spread across the whole economy. Each domain was built with at least one specialized data partner. SemiAnalysis says it was responsible for final quality assurance and was heavily involved in task iteration.

The leaderboard is tight at the top. Opus 5.5 scores 12% and Gemini 4 Carbon scores 10.9%. Chubby, a tech analyst, argued that Carbon gives up 1.1 points while costing about four times less than Opus 5.5, so it wins on cost-performance. A reply under the same post put the gap in task terms, saying Opus 5.5's lead is about two tasks out of 206. Another reply said Carbon is not publicly available, which means the cost comparison concerns a model readers may not be able to try.

The 12% figure depends on how safety classifiers are counted. SemiAnalysis said Opus 5.5 and GPT-6 Astra have had their overall scores held down by safety classifiers, which can intervene in cyber, biology and alignment tasks. It published a second set of results excluding those three domains, but its post does not give the alternative scores. In a separate analysis, Elie Bakouch finds that removing safeguard-affected benchmarks changes the picture, and he suggests OpenAI has stronger biology safeguards than Anthropic. SemiAnalysis itself says classifiers matter for model choice, claiming Muse Wonder Full is the best available cyber model by a meaningful margin.

Within-family comparisons also appear in the discussion. Carbon beats Argon across the board, and Sol 6.1 and Astra score very similarly. A separate report says some Argon cybersecurity runs may be affected by a bug in Antigravity.

Reaction has been mixed. Demis Hassabis welcomed the benchmark. The handle @dylan522p apologized to Google, calling Carbon spikier but very strong in some areas. A satirical post complained that GDPx is not open source and is costly to audit.