Best AI model scores 12% on SemiAnalysis's GDPx test

Twelve percent.

That is the best overall score any model posted on GDPx, a benchmark of long, real-work tasks that SemiAnalysis published on Nov. 2. Your agent demo looked better than that. The demo was short.

GDPx gives each task up to eight hours

GDPx covers 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis says it picked the fields where frontier models could have the most economic impact.

Each task gets a default limit of 8 hours, and some are shorter. The headline score is pass@1 across three attempts, and every domain counts equally. Finance and legal are reported as average partial credit, because SemiAnalysis says binary grading gives nearly every model a zero there.

Models ran in their own agent harness where one exists: Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon. The rest used OpenCode.

SemiAnalysis says even the best model scores just 12% overall. It calls Opus 5.5 the best all-around model, and says OpenAI and Anthropic lead while Meta, SpaceXAI and open-weight models trail. It places Gemini 4 Carbon as the only model in the same range as Opus and Astra.

No model wins everywhere. By SemiAnalysis's count, Opus 5.5 leads quant trading, hardware design and chip design. Gemini 4 Carbon leads biology and software engineering. GPT-6 Astra leads recursive self-improvement and materials science. Muse Wonder Full leads cybersecurity "by a meaningful margin."

The agent finishes the task you named and misses the rest

The failure pattern is the useful part.

On the public software task, adding Git LFS support for partial clones, none of 25 runs passed. Across its software engineering tasks, SemiAnalysis says 76% of failures involved requirements around the main request. Some were never attempted, some were misread, and some were noticed and left unfinished. It also found models tend to write self-confirming tests.

Picture a contractor who builds the deck, forgets the railing, then signs off on their own inspection.

The rules held up no better. In 200 cybersecurity runs on the alignment track, 98% broke at least one behavior rule and 87% broke one SemiAnalysis rates as severely misaligned. On a public legal task, redlining a Series B document, binary grading proved too hard for every model, though Opus 5.5 averaged 44% partial credit against 25% for GPT-6 Astra.

Yes, every AI benchmark has a method you can argue with. SemiAnalysis itself notes that safety classifiers pulled down the overall scores of Opus 5.5 and GPT-6 Astra. Still, a 12% ceiling on eight-hour work is the number to plan around.

What to do before your next long agent run

The fix is cheap.

Write the requirements down, including the ones you think are obvious. Write the acceptance checks yourself, and do not accept the agent's own tests as proof. Run your own evals on the tasks you actually hand off.

Pick the model per job, too. GDPx's domain leaders differ, and on cost against performance SemiAnalysis puts Luna and DeepSeek at the low end, Opus at the high end and OpenAI in the middle.

The next frontier launch has a harder yardstick

SemiAnalysis publishes the data at gdpx.semianalysis.com. Every model release from here can be measured on eight-hour work in public, which is harder to spin than a one-line chart in a launch post.

Watch whether the next model moves the 12%. Then check your agent's tests yourself.