Every few months a new leaderboard claims the models have finally been measured on real work. This one arrives with a number nobody can spin: the best model in the world scores 12%. SemiAnalysis released GDPx on Sunday — 206 tasks across 11 domains the firm considers where frontier intelligence will actually earn its first trillion dollars — and Anthropic's Claude Opus 5.5 leads the all-domain table at 12.0%, according to SiliconNoon.

The design is deliberately hostile. Default tasks get an eight-hour limit, models run in their native harnesses — Claude Code for Opus, Codex for OpenAI's models, Antigravity for Gemini — and the headline score is pass@1 across three attempts with every domain weighted equally. Fourteen specialist data partners built the tasks: Arithmetic on cyber, Latch Bio on biology, Living Context on finance, Chakra on legal, Third Moment on quant trading, Extrapolate on recursive self-improvement, Beautiful Machines on hardware design, ValenceForge and Emulated on materials, Fleet on alignment, Datacurve on software engineering, and Aria, Arcverex and Ergodex on chip design. Finance tasks sit in a simulated deal world seeded from a real transaction — a private equity fund buying about 62% of a snack manufacturer for $880 million plus an earnout, spread over roughly 4,000 anonymized files. The agent has to dig, not just write.

The per-domain spread is its own story. Opus 5.5 wins quant trading, hardware design, and chip design outright; Gemini 4 Carbon is the best biology model and edges software engineering; OpenAI's GPT-6 Astra wins recursive self-improvement and materials science convincingly. Finance and legal are the floor: under binary grading, "basically every model scores exactly 0." And Carbon is the number to watch — 10.9% at $16.29 per task versus Opus at 12% and $51.41, per the leaderboard at gdpx.semianalysis.com. SemiAnalysis's Dylan Patel summed it up on X: "Official apology to Google from me, they cooked with Carbon."

But the finding that will cause more fights than the 12% is the sandbagging caveat. SemiAnalysis says Opus 5.5 and GPT-6 Astra had their scores depressed by safety classifiers firing on cyber, biology, and alignment tasks — strip those three domains and the ranking moves. Its conclusion is that the best cyber model available today is Meta's Muse Wonder Full "by a meaningful margin," because it refuses less. Any capability comparison between models with different refusal policies is quietly comparing policy as much as skill, and that distinction vanishes the moment the number is printed on a launch slide.

Two caveats worth holding before quoting the numbers: GDPx is a paid SemiAnalysis product with an interest in being the benchmark of record, and no outside group has replicated these results. What to watch next: whether labs start reporting GDPx alongside GDPval on model cards, and whether the task set leaks into the next round of training data — the two ways a hard benchmark quietly stops being hard.

— via SiliconNoon