📅 2 November 2026 • SemiAnalysis / Unite AI
SemiAnalysis has released GDPx v1.0, an agentic benchmark designed to measure frontier models on real professional work rather than isolated quiz-style questions. The benchmark runs 206 tasks inside live environments across 11 professional domains, ranging from chip design to quantitative trading. Task data and environments come from 14 partner organizations that supply the material practitioners actually work with, and each task carries a default budget of 8 hours. The current leaderboard ranks 13 models.
Claude Opus 5.5 leads GDPx with a score of 12.0%. Gemini 4 Carbon is second at 10.9%, spending $16.29 per task on average. GPT-6 Astra takes third place with 9.6% at a markedly higher $53.99 per task. The leader is not cheap either: Claude Opus 5.5 averages $51.41 and consumes about 548,000 output tokens per task. That the best score on the leaderboard is just 12% indicates how much headroom remains on long-horizon professional work, where a single run can take hours and produce thousands of intermediate steps.
The grading detail is as revealing as the ranking. Under strict binary evaluation, no model solved a single task in the finance or law domains, despite these fields being rich in structured, rules-based work. In quantitative trading, 70 of 111 graded runs landed below the reference strategy, meaning most attempts lost value relative to a baseline a professional would be expected to beat. The pattern suggests models can produce plausible intermediate work while still failing the end-to-end outcome that clients actually pay for.
SemiAnalysis positions GDPx as the successor to OpenAI GDPval and plans to ship new versions every 3-6 months. The cadence matters because benchmarks that rank models on durable professional tasks tend to age quickly as labs optimize against them. Each new version, live environments and partner-supplied data are meant to keep the measurement current and harder to game. The first release also sets a baseline the industry can track as next-generation models arrive.
For teams selecting models for agentic deployments, GDPx adds a cost-aware view of capability that headline scores usually hide. The gap between Gemini 4 Carbon and GPT-6 Astra on cost per task, set against a roughly 1.3-point score difference, shows how differently frontier models trade accuracy against budget. Automation programs can now weigh per-task economics alongside capability scores when mapping which workflows to hand to which model.
IMPACT: Evaluation teams, platform engineers and anyone buying model capacity for agentic workflows are directly affected. Model selection can now weigh per-task cost against pass rates on professional work instead of headline scores, and the zero-solve results in finance and law mark those domains as not ready for unsupervised delegation. Teams running quant or legal pilots should add reference-strategy comparisons to internal evals. What to watch: whether leader scores climb with each 3-6 month release, how cost per task shifts as labs optimize, and which domains move off zero solves first.
Sources: newsletter.semianalysis.com, unite.ai
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)