Claude Opus 5.5 Tops Semi Analysis’ GDPx Benchmark, Gemini 4 Carbon Placed Second

Semi Analysis released GDPx, a new 11-domain, 206-task benchmark for agentic work in high-revenue industries, and Anthropic's Claude Opus 5.5 topped it with a score of just 12%, ahead of Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%. Semi Analysis reported that Claude Opus 5.5 costs $51.41 per task, roughly three times Gemini 4 Carbon's price for a score barely a point higher, while GPT-6 Astra is the priciest at $53.99 per task. On a second leaderboard excluding cybersecurity, biology and alignment, GPT-6 Astra edges Claude Opus 5.5 by a tenth of a percentage point, which Semi Analysis said aligns more closely with its own subjective sense of model performance.

Semi Analysis, the semiconductor and AI research outfit, has released a new benchmark called GDPx, built specifically to measure how well frontier AI models perform agentic work in the handful of industries the firm believes will actually drive AI revenue. Anthropic’s Claude Opus 5.5 comes out on top with a score of just 12%, followed by Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%. The low scores are the point. Semi Analysis is positioning GDPx as a genuinely frontier-level test, one where even the best model in the world clears roughly one out of every seven tasks thrown at it. That’s a deliberate departure from the benchmarks that have increasingly been called into question for being easy to game or too narrow to mean much, a problem Semi Analysis itself has flagged before when it called out models for being over-optimized for specific benchmark results https://officechai.com/ai/gemini-3-8-flash-muse-spark-1-3-most-benchmaxxed-models-semi-analysis/ rather than genuine capability gains. Why Semi Analysis built GDPx this way The firm’s reasoning for the benchmark’s scope is laid out plainly in its announcement thread: it doesn’t believe AI’s economic impact will spread evenly across the economy. Instead, Semi Analysis expects the first trillion dollars of AI revenue to concentrate in a small number of sectors where frontier intelligence can be deployed at high return on investment. Building a benchmark around GDP share or standard occupational classifications like SOC or O NET, the firm argues, would miss that concentration entirely. That’s why GDPx skips categories like government, real estate, and insurance in favor of chip design, cybersecurity, and biology, domains Semi Analysis expects frontier models to reshape first. Measuring progress on the benchmark is a collaborative effort too. Across its 11 domains, GDPx draws on 206 tasks built with at least one specialized data partner per domain, including names like Aria, Arcverex, and Ergodex for chip design, Living Context for finance, and LatchBio for biology, with Semi Analysis handling final QA and task iteration throughout. The leaderboard Here’s how the field stacks up across all 11 domains: Claude Opus 5.5 leads comfortably, but it isn’t the cheapest or the fastest way to get there. At $51.41 per task, it costs roughly three times what Gemini 4 Carbon https://officechai.com/ai/gemini-4-carbon-benchmarks/ charges for a score that’s barely more than a point higher. GPT-6 Astra, meanwhile, is the priciest model on the list at $53.99 per task, for a score that lands in third. Safety classifiers skew the overall numbers Semi Analysis makes a specific point of flagging that two of the top models, Claude Opus 5.5 and GPT-6 Astra, are held back on the overall leaderboard by their own safety classifiers, which kick in and affect performance in domains like cybersecurity and biology. To account for this, the firm published a second leaderboard that strips out cybersecurity, biology, and alignment entirely, leaving eight domains: With those three domains excluded, GPT-6 Astra edges ahead of Claude Opus 5.5 by a tenth of a percentage point, and Gemini 4 Carbon slips from second to fifth. Semi Analysis noted that this adjusted ranking lines up more closely with the firm’s own subjective sense of how the models actually perform, which it offered as a kind of informal validation of the benchmark’s design. The bigger picture What stands out most in Semi Analysis’ release isn’t really the ordering of any one leaderboard, it’s the headline number. A 12% score at the very top of the field, across tasks drawn from domains the firm considers the most commercially consequential for AI, is a reminder of how much distance remains between today’s frontier models and anything resembling full automation of high-value knowledge work.