Opus 5.5 GDPx Score of 12% Explained

SemiAnalysis launched GDPx with Claude Opus 5.5 in first place at 12.0%, ahead of Gemini 4 Carbon at 10.9%. In the same article, the authors write that certain models, Opus 5.5 among them, have "sandbagged overall scores due to their safety classifiers." A leaderboard whose authors say it understates the winner is hard to compress into a headline.

Verdict: true. The GDPx leaderboard lists Opus 5.5 first at 12.0% and Gemini 4 Carbon second at 10.9%.

The 12.0% averages unlike things

Start with what the number is. GDPx scores each model across 11 domains, weights the domains equally, and weights each partner inside a domain equally. The benchmark holds 206 tasks from 14 data partners, so the 10 tasks in RSI carry the same weight as the 34 in chip design.

Grading mixes binary and continuous scoring depending on the domain, and the default score averages three attempts. The 12.0% is therefore not a share of tasks completed. The authors call GDPx "a true frontier-level benchmark," so a low top score is intended headroom.

Finance and legal distort the arithmetic in a specific way. The authors say every model scores 0% on finance under binary grading, and legal goes the same way, so they report average partial scores for both. The excerpt does not say whether those partial scores feed the headline figure. The leaderboard has a separate Partial toggle, and my reading is that it exists because the choice matters.

The domains also reflect a view. The authors argue the first trillion dollars of AI revenue will concentrate in sectors with high ROI, and they left out government, real estate and insurance. I infer a different selection would reorder the table.

Each model ran in a different tool

Opus 5.5 ran in Claude Code, GPT-6 Astra in Codex and Gemini 4 Carbon in Antigravity. Models without a tool of their own ran in OpenCode, and the default limit was 8 hours per task. I infer the score mixes model quality with tool quality, and nothing in the excerpt separates the two.

The excerpt also carries no confidence intervals, no spread across the three attempts and no count of runs. With 12.0 against 10.9, you have no way to tell whether the order survives a rerun.

Two of the authors, Andrew Megalaa and Max Kan, previously co-founded an RL environment startup which sold data to frontier labs. I have no evidence of bias in the numbers. The excerpt documents no independent replication either.

When "sandbagged" means Opus 5.5's 12% is the conservative number

The authors then present a second ranking without cyber, biology and alignment. The excerpt elides it. This is the best case against calling the headline overstated. If classifiers hold Opus 5.5 down in three domains, 12.0% reads low, and asking for variance data means asking for precision on a number the authors already distrust. They call Opus 5.5 "the best all-around model" and credit it with quant trading, hardware design and chip design. Carbon, they write, is "the only model in the same ballpark as Opus and Astra."

I think the objection holds on direction and fails on size, because the excerpt measures no size. One secondary write-up says the order "tightens noticeably" without the three domains. Another calls the rankings "materially different" and relays another outlet's view that Astra ranks higher on the domains it attempts. Neither says Opus 5.5's lead grows.

Carbon wins biology, one of the three excluded domains. Two secondary write-ups say Meta's Muse Wonder Full tops cyber. The excerpt does not show Opus 5.5 leading any of the three. If classifiers move scores, 12.0% measures capability and classifier behavior together, and whether removing them helps Opus 5.5 or Astra more depends on a ranking the excerpt withholds.

The cheaper number sits one row down

Opus 5.5 costs $51.41 per task. Carbon costs $16.29 and scores 10.9%. GPT-6 Astra costs $53.99 and scores 9.6%. Meta's Muse Wonder Full is fourth at 9.0% with no public price, so it never appears on the cost chart.

On 8-hour tasks where the best model reaches 12.0%, a buyer might reasonably pay for the top score anyway. The authors say they will want impressions once Carbon becomes generally available. I infer it was not on sale to everyone when tested, so $16.29 prices a model not yet on general sale.

Where to look before you repeat the rank

The informative parts of GDPx are the caveats the authors published themselves. The one figure to settle them, the second ranking, sits in an article marked paid, with its excerpt heavily elided. The leaderboard lists 13 models of a larger set of 20, and the excerpt never says what the other 7 score.

SemiAnalysis plans a new version every 3-6 months. The first thing to look for in it is a spread across the three attempts.