On 2 November, the research firm SemiAnalysis published GDPx, a new benchmark built to measure how well AI models handle long, realistic work in the fields where it thinks frontier AI will matter most. The headline result is that even the best model scores 12%. In two of the fields, finance and legal, SemiAnalysis says nearly every model scores exactly zero.
What GDPx measures
SemiAnalysis covers 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (AI systems that help build better AI), hardware design, materials science, alignment and software engineering. It says it deliberately left out industries such as government, real estate and insurance, because its goal is to track the areas where it expects models to have the most economic impact.
The tasks were built with 14 specialist data partners, mostly companies that build training environments for AI. OfficeChai, which read the published leaderboard, reports 206 tasks in total. Each task runs in a realistic setting with real tools and files. One finance task, for example, starts from a real private equity deal and about 4,000 anonymized files. The default time limit is 8 hours, and models run in their makers' own tools where available, such as Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon. The headline score is pass or fail on the first try, averaged across three attempts, with every domain weighted equally.
The leaderboard, field by field
On the overall leaderboard, as reported by OfficeChai and Crypto Briefing, Claude Opus 5.5 scores 12.0%, Gemini 4 Carbon 10.9% and GPT-6 Astra 9.6%. SemiAnalysis does not crown a single winner everywhere. In its write-up:
- Claude Opus 5.5 is described as the best all-round model, with wins in quant trading, hardware design and chip design.
- Gemini 4 Carbon is described as the best biology model, and it narrowly wins software engineering.
- GPT-6 Astra wins recursive self-improvement and is the best model for materials science.
- Muse Wonder Full is named the best cybersecurity model by a meaningful margin, largely because other models' safety classifiers hold them back.
Those safety classifiers matter for the overall ranking. SemiAnalysis says the scores of Opus 5.5 and GPT-6 Astra are pulled down because their safety systems block some work in cybersecurity and biology. According to OfficeChai, when those areas and alignment are removed, GPT-6 Astra edges ahead of Opus 5.5 by 0.1 percentage points and Gemini 4 Carbon falls to fifth. The gaps at the top are small enough that the order could change with the next version. On cost, OfficeChai reports that Opus 5.5 averages $51.41 per task, and SemiAnalysis says cheaper models such as Luna and DeepSeek compete at the low end, OpenAI takes the middle and Opus wins the high end.
The zero scores in finance and legal
The most striking line in the report is about finance and legal. SemiAnalysis writes that its tasks in those two fields are so hard that nearly every model scores exactly zero under pass or fail grading, so for legal it also reports partial-credit scores. These are long, messy tasks built from real material, which is very different from answering exam questions or drafting a single clause.
The alignment results tell a related story. In cybersecurity scenarios supplied by the data partner Fleet, 98% of runs broke at least one behaviour rule set for the task, which suggests that models working alone for hours still need rules enforced from outside.
Limits worth knowing
GDPx is new and has not been checked by anyone outside SemiAnalysis. Most of the report and the full score tables sit behind a paywall, and the leaderboard is published as images, so the overall figures above come from outlets that read those images. The task partners are data companies that sell training environments to AI labs, and SemiAnalysis says it plans to showcase the best of those vendors, so the benchmark is also a shop window for them. The authors say the results line up with their own impression of the models, which is a judgment rather than independent validation. SemiAnalysis plans a new version every three to six months, so treat today's numbers as a first reading.
What this means for jobs
A lot of the public conversation about AI and work swings between two extremes. GDPx is a useful middle point. Today's best models can clearly help with parts of expert work, and they still fail most of the time when handed a full day of it with real files and no human steering. That gap is where most professionals actually spend their time. We think this is a more honest guide to near-term job risk than any single demo, while also being a reminder that benchmarks like this get updated as models improve.
For a lawyer or a finance analyst, the zero scores mean AI remains a drafting and checking aid for full deal or case work and cannot yet take that work over. For an engineering lead, the higher scores in chip design and software suggest it is worth giving agents bounded, well-specified pieces of a project while a person owns the whole.
What to do this week
Test your own job the GDPx way. Pick one task that takes you most of a day, give your AI tool the real files you would use, judge the finished result as pass or fail, and write down the exact step where it breaks. That step is a good description of the skill that still makes you valuable. If you want help choosing which assistant to test, start with our comparison page or the Claude guide in Learn.