Vibeus12.0% for Claude Opus 5.5, 10.9% for Gemini 4 Carbon and 9.6% for... - Vibeus
12.0% for Claude Opus 5.5, 10.9% for Gemini 4 Carbon and 9.6% for GPT-6 Astra: the top three in the all-domain GDPx snapshot. Muse Wonder Full follows at 9.0%, Claude Fable 5.1 at 8.5% and GPT 6.1 Sol at 8.4%. Both attached charts show a selected 13 of the 20 models, not the whole model list.
SemiAnalysis’s new benchmark spans chip design, cybersecurity, biology, finance, legal work, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Its premise is that the first trillion dollars of AI revenue will concentrate in sectors where frontier intelligence can deliver high returns. That is the researchers’ argument for choosing these domains instead of weighting a benchmark by GDP share or the SOC/O*NET occupational taxonomy; government, real estate and insurance are not standalone domains. Alignment tasks can still use settings such as insurance and healthcare.
The software-engineering view tells a different story. Across 15 tasks involving long-horizon feature work, debugging and migrations in large production codebases, Gemini 4 Carbon scores 11.1%. Opus 5.5, GPT 6.1 Sol and Gemini 4 Argon each score 8.9%; GPT-6 Astra scores 6.7%. Reported cost per task is $8.16 for Carbon, $16.97 for Opus and $2.57 for Sol in this view. These are benchmark run costs, not subscription prices.
The percentages need their grading context. GDPx combines binary and continuous grading depending on the domain, averages each task over three attempts, and weights the 11 domains equally, with equal weight for partners inside a domain. A 12% aggregate score is therefore not simply the fraction of all tasks solved.
SemiAnalysis also says safety classifiers depress some models’ results in cybersecurity, biology and alignment; excluding those domains changes the ranking. The authors present GDPx as a frontier-level alternative to GDPval, built around demanding agent tasks and specialized environments. The scores and cost columns belong to this evaluation setup, rather than a universal ranking of model quality.
