• On the GDPx v1.0 leaderboard, the top score is 12.0 percent (Claude Opus 5.5), and 13 models are ranked in total.
• Gemini 4 Carbon scores 10.9 percent at $16.29 per task, against $51.41 per task for the leader, so cost per point differs sharply.
• The suite covers 206 tasks in 11 domains, each weighted equally, with every task run three times and averaged.
• Critics cite thin sub-domain samples, LLM-based rubric grading in some domains, and different agent harnesses per model.
• Signal, with caveats: the absolute gap to useful autonomy is credible, but the ordering at the top is not.

Signal. SemiAnalysis' GDPx, released in early November 2026, finds that no frontier AI agent completes more than 12 percent of its simulated professional work, with Claude Opus 5.5 on top at 12.0 percent. The benchmark spans 206 tasks in 11 domains, from chip design to quant trading. The absolute level is the story: the ordering among the top three is within what small samples can blur, but the distance from reliable autonomous work is not.

What the leaderboard actually shows

The GDPx v1.0 leaderboard ranks 13 models, led by Claude Opus 5.5 at 12.0 percent and $51.41 per task. Gemini 4 Carbon follows at 10.9 percent and $16.29, and GPT-6 Astra at 9.6 percent and $53.99. At the low end, GPT-6 Luna and DeepSeek 4.1 Flash 1017 score 5.1 and 5.4 percent at under $1 per task. SemiAnalysis lists output tokens too, so cost and verbosity can be compared directly.

How the score is built

The headline number is an equal-weighted average over 11 domains, and over partners within each domain, using each task's intended grading method and three attempts. SemiAnalysis says its default limit is eight hours, and that models run in their native harness where one exists, such as Claude Code for Opus 5.5. Finance and legal use partial credit because, The Decoder reports, strict pass/fail would leave no model with a completed task.

Where the ranking is weakest

The top-three ordering is fragile. Developer Josh Purtell, quoted by The Decoder, calls GDPx more noise than signal, citing sub-domains with about ten tasks and per-provider harnesses that favor stronger tooling. The Decoder also reports that law and hardware design are graded against rubrics by language models, and that one example task per partner is usually public, so outsiders cannot fully audit difficulty. Keeping tasks secret does limit training on the test.

Safety filters change the picture

According to The Decoder, removing the cyber, biology and alignment domains lets GPT-6 Astra edge past Claude Opus 5.5 at 11.8 percent. It also reports GPT-6 Astra and GPT 6.1 Sol scoring 0 percent in cyber while Muse Wonder Full reaches 35.2 percent. We could not verify these figures against the leaderboard page, so treat them as reported. The practical point: refusal policy is part of the score.

Verdict for teams shipping agents

Use GDPx as evidence that long-horizon professional tasks remain largely unsolved, not as a model-selection tool. The 1.1-point gap between first and second is small next to the roughly threefold cost difference, so a production team should rerun its own tasks, with its own harness, before paying for the leader. SemiAnalysis plans updates every three to six months, and the next version should show whether the sample-size objections get addressed.