Claude Opus 5.5 just topped the toughest benchmark frontier models have ever faced. Its score was 12.0%. Gemini 4 Carbon managed 10.9%. GPT-6 Astra, 9.6%.
On the tasks where analysts expect the first trillion dollars of AI revenue to actually appear, the best machine on the planet gets almost nine in ten questions wrong.
On Nov 2, SemiAnalysis launched GDPx, a benchmark covering 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. The selection logic is blunt. Rather than measuring models across the whole economy, GDPx measures them where deployable ROI is highest — the fields where AI is most likely to earn serious money first.
That framing is the whole point. The benchmarks everyone cites are saturated; when every frontier model scores above 90%, a benchmark stops telling you anything. SemiAnalysis built one where saturation is years away. The leaderboard above is the proof.
Finance and legal scored a flat zero
Not low. Zero. Every model tested scored 0% on finance tasks and 0% on legal tasks.
If your 2027 plan involves an AI analyst or an AI paralegal doing unsupervised work, this is the number to sit with. The current frontier cannot do that job at a level the benchmark can even detect. Whatever AI is doing in those industries today, it is assistance and drafting work — not autonomous professional work.
The leaderboard is partly a safety filter
The finding that interests me most isn’t a headline score. SemiAnalysis noticed that safety classifiers visibly depress scores in cybersecurity, biology, and alignment. Exclude those three domains from the scoring and the ranking order shifts.
Think through what that means. When your company shortlists models for sensitive work, the list is being sorted partly by what labs allow the models to do, not purely by what the models can do. Muse Wonder Full is reportedly the best cybersecurity model available right now, and reportedly for exactly this reason — its answers pass through a less aggressive filter, so more of its real capability shows up on the board.
Capability comparisons in sensitive domains are, right now, also filter comparisons. Most procurement processes do not account for this at all.
There is no best model
The GDPx results are spiky, and the spikes do not overlap:
- Claude Opus 5.5 leads in quant trading, hardware design, and chip design
- Gemini 4 Carbon takes biology and software engineering
- GPT-6 Astra wins recursive self-improvement and materials science
Nobody wins everywhere. There is no general-purpose champion, and the gaps between leaders and losers within a single domain are large enough to change what you should buy.
I think that’s a healthier situation than one dominant model. It forces an honest question that aggregate leaderboards let you avoid: best at what?
What to do with a 12% frontier
Stop buying on aggregate scores. If your vendor pitch cites a single overall number, ask for the domain breakdown. GDPx exists precisely because the aggregate hides everything important. Build your own eval on your own tasks — a half-day of domain-specific testing beats any leaderboard.
Treat finance and legal AI as R&D, not deployment. Zero is zero. Prototype and measure — but keep a human signatory on everything until a benchmark like this says otherwise.
Audit your sensitive-domain stack for the filter effect. If you work in cyber, bio, or alignment-adjacent fields, test at least one model you already dismissed. The refusal layer may be the reason it underperformed, not the underlying capability.
Re-run your evals every quarter. GDPx refreshes every 3 to 6 months with input from 14 specialized data partners. Scores will move. A model choice made in November 2026 has a shelf life, and it is shorter than your annual procurement cycle.
Reset internal expectations. If a stakeholder asks why the AI pilot “only” works on half the tasks, this benchmark is your answer. The ceiling for the best model in the world, on the highest-value work in the economy, is 12%. Everything shipped today is an experiment against that ceiling.
The low ceiling is the good news
A 12% ceiling reads as disappointment. It shouldn’t.
If frontier models were already near-competent at chip design and quant trading, the next decade would be a deployment story — incumbents bolting on finished intelligence and cutting headcount. They aren’t near-competent. There is a decade of buildout in front of everyone, and the gaps between models mean no single vendor locks up the market before you’ve picked your stack.
SemiAnalysis plans to refresh GDPx every few months. The first number to watch is not who leads at 12%. It is whether finance and legal stay at zero. The moment one of them blinks, a trillion-dollar industry quietly changes shape.
Based on SemiAnalysis’s GDPx launch, published Nov 2, 2026.
