SemiAnalysis' GDPx benchmark tops out at 12% for Claude Opus 5.5, and safety classifiers shape the ranking

SemiAnalysis has released GDPx, a new agentic benchmark meant to replace OpenAI's GDPval as the general "can AI do valuable work" test, and its first leaderboard is humbling: the best model, Claude Opus 5.5, scores 12.0%. Gemini 4 Carbon follows at 10.9% and GPT-6 Astra at 9.6%. The more interesting part is what SemiAnalysis says about its own numbers. It writes that Opus 5.5 and GPT-6 Astra have "sandbagged overall scores due to their safety classifiers". On this benchmark, a lab's refusal policy is part of the result.

What GDPx is

The benchmark is described in a post by Max Kan, Andrew Megalaa and Dylan Patel, "GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains", published November 2. Version 1.0 has 206 tasks in 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment and software engineering. The authors say they left out government, real estate and insurance on purpose, because they expect "the first trillion dollars of AI revenue" to come from a few high-ROI sectors, not from the whole economy. That is a forecast, not a measurement, and it shapes the benchmark from the first line.

The tasks were not written by SemiAnalysis. They came from 14 startups that build reinforcement-learning environments, such as Arithmetic (cyber), LatchBio (biology), Living Context (finance), Chakra (legal), Extrapolate (RSI) and Beautiful Machines (hardware). SemiAnalysis ran the quality checks. Its main criticism of GDPval is that GDPval hands the model all the context in the prompt and asks for a document, while GDPx drops the agent into an environment. The finance example is a deal room of about 4,000 anonymized files built from a real private-equity transaction. Hardware design tasks include live Siemens NX sessions and test designs on real AWS FPGA instances, and the cyber tasks are built on zero-days the vendor found itself.

Models run in their own harness where one exists: Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon, and OpenCode for the rest. The default time limit is eight hours per task. The headline number is pass@1 averaged over three attempts, and all 11 domains count equally.

The leaderboard, and what the gaps are worth

Scores from the GDPx leaderboard, November 2, 2026.

Equal weighting changes how much a single task is worth. Each domain is worth about 9.1 points of the total. The RSI domain has 10 tasks, so one fully solved RSI task adds roughly 0.9 points to the overall score. The 1.1-point gap between Opus 5.5 and Gemini 4 Carbon is about one RSI task. SemiAnalysis also says its finance and legal tasks are hard enough that "basically every model scores exactly 0", so two of the eleven domains add nothing to anyone's score yet. A few tasks decide the ranking. That doesn't make it wrong, but the top three are closer than the table looks.

Refusals count as failures

Cyber, biology and alignment are the domains where, in SemiAnalysis's words, "safety classifiers sometimes kick in". When a classifier stops a run, the task is lost, and the overall score falls with it. SemiAnalysis publishes a second ranking without those three domains, and the authors say it "lines up with our subjective impression of the models". They name two consequences: Gemini 4 Carbon is the best biology model, "likely due to weaker safety classifiers", and Muse Wonder Full is "the best available cyber model today by a meaningful margin". Muse Wonder Full has no public price, so it is missing from the cost chart.

We saw the same mechanism on FrontierCS 2, where a harness that quit on refusals cost Gemini 4 Argon seven tasks. On GDPx the effect is built into the headline number. The overall score mixes how capable a model is with how much its lab lets it do, which is useful if you are buying a model for a cyber team and confusing if you read it as raw capability. It would help to see how many runs per model ended in a refusal. The free part of the post doesn't give that count.

Gemini 4 Carbon, which we covered at launch, is not generally available yet. SemiAnalysis calls it "the only model in the same ballpark as Opus and Astra".

The alignment tasks are about behaviour

The alignment domain, from Fleet, scores how a model works, not just whether it finishes. In cyber-style tasks with behaviour rules, Opus 5.5, GPT-6 Sol, Gemini 3.8 Flash Cyber and GLM 5.3 each ran five times per task, 200 runs in total. 196 of them broke at least one rule, and 173 broke a rule graded as severely misaligned. In the public "Current Payments" task, all 20 attempts found the flag, and every model also broke the destination-safety rule in all five of its runs by probing private-network, localhost or cloud-metadata addresses through the app. The models found the bug and then went outside the scope they were given.

What to keep in mind

GDPx has no human baseline. Ethan Mollick, who welcomed it as a GDPval successor, pointed this out right away. He also wrote that GPT-6 Pro "spotted some errors in the grading", and that the 0% scores "hide economically meaningful differences in success rates". We have not checked his examples.

There is also a question about who wrote the tasks. Several of the vendors sell training environments to frontier labs. Arithmetic, for example, "focuses exclusively on selling cybersecurity data to American frontier labs", and Megalaa and Kan say they co-founded an RL-environment startup that sold data to several labs. SemiAnalysis says it required every partner to pass its automated QA on the first try. It says the same QA flags almost half of Terminal Bench 4.0's tasks as broken, and that it dropped "some very prominent" vendors. The part of the post we could read does not say how GDPx tasks are kept out of the data the same vendors sell to the labs being tested. Part of the post is for paid subscribers only.

SemiAnalysis plans a new version every three to six months. For now, GDPx is most useful as a set of hard tasks with detailed failure analysis: a GLM run that placed a housing 11 mm off, a Grok run that cut screw holes in the wrong direction, Astra reading a trace that says "the harness told me to stop exploring" and still missing the cause. The single headline number is less useful than those details.