The GDPx benchmark is a new test of AI agents from the research firm SemiAnalysis, released on November 2, 2026. Version 1.0 has 206 tasks across 11 domains (chip design, cybersecurity, biology, finance, legal, quant trading, recursive self improvement, hardware design, materials science, alignment and software engineering), built with 14 specialist data companies. Each task drops the model into a realistic work environment with up to 8 hours to finish. The best model, Claude Opus 5.5, scores 12.0%, so GDPx is far from saturated and small score gaps carry little meaning.
This guide explains how the GDPx leaderboard is built, how the headline score is calculated, what the first results say, and the limits you should keep in mind before quoting a GDPx number. All figures come from the official GDPx site (data dated November 1, 2026) and the SemiAnalysis launch post, both read on the morning of November 2. We picked no "best model" here; the point is to help you read the board.
What the GDPx benchmark measures
GDPx sets out to measure general agentic capability: whether a model, working inside its own agent harness, can finish long, open ended professional work in the fields SemiAnalysis expects AI to change most in the next one to two years. The authors (Max Kan, Andrew Megalaa and Dylan Patel) say they chose domains by expected economic impact for frontier models, not by share of GDP or by an occupation list, and they skipped fields such as government, real estate and insurance on purpose.
The name invites a comparison with OpenAI's GDPval benchmark (paper), and SemiAnalysis makes that comparison itself. Its main complaint is that GDPval tasks hand the model all the context in the prompt and ask for a one shot deliverable, while real work starts with digging through files, systems and email. GDPx tasks are built around that digging. One finance environment, for example, is a replica of a private equity deal with roughly 4,000 anonymized files; the model has to find the right documents before it can do the accounting.
The 11 domains and 206 tasks
The domains page lists each domain, its partner and its task count:
| Domain | Partner(s) | Tasks | What the tasks look like |
|---|---|---|---|
| Chip design | Aria, Arcverex, Ergodex | 34 | Verification, RTL design under power, performance and area limits, physical design |
| Finance | Living Context | 24 | Deal modeling and transaction accounting in a replicated deal room |
| Materials science | ValenceForge, Emulated | 23 | Driving simulated lab equipment and reading raw experimental data |
| Hardware design | Beautiful Machines | 21 | CAD, CAM, PCB layout and FPGA work in real engineering software |
| Legal | Chakra | 20 | Research, drafting, redlining and issue spotting |
| Cybersecurity | Arithmetic | 18 | Finding, exploiting and fixing private zero days in live services |
| Alignment | Fleet | 18 | Whether agents stay honest and inside their authority under pressure |
| Software engineering | Datacurve | 15 | Feature work, debugging and migrations in large codebases |
| Biology | LatchBio | 13 | Analysis and decisions on real experimental data |
| Recursive self improvement | Extrapolate | 10 | ML research engineering, such as debugging a broken RL training run |
| Quant trading | Third Moment | 10 | Finding trading edge that holds on hidden days of order book data |
Chip design has three partners and the most tasks; SemiAnalysis openly says it wanted chip design slightly overrepresented. Alignment is the odd one out because it measures behavior rather than skill: the cyber tasks, for instance, check whether an agent stays inside its rules of engagement while it hunts for an exploit, which overlaps with the work covered in our LLM red teaming tools guide.
How the GDPx score is calculated
The scoring rules, from the launch post and the site:
- Attempts: each model gets three attempts per task, and the headline is pass@1 averaged over those three runs.
- Grading: each domain uses its intended method, either binary (pass or fail against a threshold) or continuous partial credit. Graders range from deterministic checks (finance spreadsheets, quant trading P&L tests, chip sign off tools) to expert rubrics judged by models, which is where the usual LLM as a judge caveats apply.
- Weights: all 11 domains count equally, and where a domain has several partners, each partner counts equally inside it. A 10 task domain therefore weighs as much as the 34 task chip design domain.
- Time and harness: the default limit is 8 hours per task. Models run in their own native harness where one exists (Claude Code for Claude models, Codex for OpenAI models, Antigravity for Gemini), and in OpenCode otherwise.
- Privacy: each partner published the full rollouts for one public task on the rollouts page; every other task stays private to stop labs from training toward it.
Two settings change what you see on the site. The default board shows 13 of the 20 models tested. A partial score toggle swaps binary results for partial credit, which matters a lot in finance and legal, where binary scores are almost all zero.
First results: who leads GDPx
The default board on November 2, 2026, with cost per task from the site's published data:
| Model | GDPx score | Average cost per task |
|---|---|---|
| Claude Opus 5.5 | 12.0% | $51.41 |
| Gemini 4 Carbon | 10.9% | $16.29 |
| GPT-6 Astra | 9.6% | $53.99 |
| Muse Wonder Full | 9.0% | no public price |
| Claude Fable 5.1 | 8.5% | $67.16 |
| GPT 6.1 Sol | 8.4% | $9.80 |
| Claude Sonnet 5.5 | 8.2% | $45.27 |
| Gemini 4 Argon | 7.3% | $12.79 |
| GLM 5.3 (lowest shown) | 2.7% | $11.79 |
Among the seven models hidden by default, the cheapest are GPT-6 Luna at $0.92 a task (5.1%) and DeepSeek 4.1 Flash 1017 at $0.94 (5.4%). The lowest of all 20 in the data is Muse Spark Open at about 1.7%. Token use varies as much as cost: Opus 5.5 averaged about 548,000 output tokens per task against about 275,000 for GPT-6 Astra.
Safety classifiers move the ranking
SemiAnalysis says safety classifiers held back some models in cybersecurity, biology and alignment, and the site has a view that leaves those three domains out. Averaging the remaining eight domains from the site's data, GPT-6 Astra (about 11.8) edges Opus 5.5 (about 11.7), GPT 6.1 Sol rises to about 10.6, and Gemini 4 Carbon falls to about 9.2. Meta's Muse Wonder Full shows the reverse effect: it tops cybersecurity at about 35%, which drops sharply once that domain is removed. A model that refuses a task scores zero whether or not it could have done it, so this is a policy result as much as a skill result.
Domain leaders are spiky
Reading the per domain results across all 20 models: Muse Wonder Full leads cybersecurity and alignment, Gemini 4 Carbon leads biology (about 25.6%) and software engineering (about 11.1%), GPT-6 Astra leads recursive self improvement (about 24.9%) and materials science (about 21.6%), and Opus 5.5 leads hardware design (about 18.7%). Two spots disagree with the launch post's prose. It names Opus 5.5 the quant trading winner, while the site data has Claude Fable 5.1 slightly ahead (28.6 to 27.4), and GPT 6.1 Sol has the top chip design average in the data. Check the live domain view before you quote a domain winner.
Finance and legal are near zero
Under binary grading, finance and legal are almost empty: on legal every model scores 0, and on finance only Gemini 4 Carbon gets above zero (about 2.8%). With partial credit, Opus 5.5 averages about 55% on finance and the top models land near 30% on legal. The public examples show why. On Living Context's quality of earnings task, no model solved any of 25 attempts, and the best single run passed 8 of 11 checks. On Chakra's Series B redline task, none of 20 rollouts produced a usable redline.
GDPx vs GDPval, Artificial Analysis and Fireworks SII
| Item | GDPx | GDPval | AA Intelligence Index v5 | Fireworks SII |
|---|---|---|---|---|
| Publisher | SemiAnalysis | OpenAI | Artificial Analysis | Fireworks AI |
| Task style | Agentic, in replicated work environments, up to 8 hours | One shot deliverables from a prompt and files | Nine evaluations, mixed formats | Partner benchmarks, each in its own harness |
| Domains | 11 chosen by expected AI impact | 44 occupations in 9 GDP sectors | Knowledge work, coding, science, general | 14 domains |
| Overall score | Yes, equal domain weights | Win rate against experts | Yes, 25% per group | No composite |
| Harness | Each lab's own agent where available | OpenAI setup (GDPval-AA uses its own) | Artificial Analysis harness | One harness per benchmark |
| Open tasks | One public task per partner | 220 public tasks | Mix of public and private | Varies by board |
The short version: GDPx is closest to a stress test of finished agent products, since Claude Code, Codex and Antigravity differ in more than the model. The Artificial Analysis Intelligence Index v5 holds the harness steadier and covers more models (methodology), while the Fireworks Specialized Intelligence Index refuses to blend domains at all (methodology). Our LLM leaderboards guide covers the wider set.
Limits to keep in mind before you quote GDPx
- Few tasks per domain. Domains hold 10 to 34 tasks. By our arithmetic, one task in a 10 task domain moving from fail to full pass shifts that domain by 10 points and the headline by about 0.9 points. Gaps of one point near the top are within that swing.
- No published error bars. We found no confidence intervals on the board. Three attempts per task reduce noise but do not remove it.
- Harness and model are mixed. A native harness result tests the agent product. Models without one run in OpenCode, which may help or hurt them.
- Mostly private tasks. You can read one public task per partner, but you cannot rerun the other tasks yourself.
- Interested parties. The partners sell training data and environments to labs and are showcased by the benchmark; SemiAnalysis also sells research on AI demand, and Fireworks served the open models. None of this makes the numbers wrong, but it is worth knowing.
- Self reported QA. SemiAnalysis says it contacted 30+ RL environment companies, required partners to pass its automated QA on the first try, and that the same QA flags nearly half of Terminal-Bench 4.0 tasks as broken. Those are its own claims, not an independent audit.
- A moving target. SemiAnalysis plans a new version every 3 to 6 months, so always cite the version (1.0) and the date.
If you want to know which model suits your own workload, treat GDPx as one input and run a small evaluation on your own tasks; our guide to how to evaluate LLM model performance walks through that.
FAQ
What is the GDPx benchmark?
GDPx is an agentic AI benchmark from SemiAnalysis, released November 2, 2026. Version 1.0 has 206 tasks across 11 professional domains, built with 14 specialist data partners, and runs each model in a realistic work environment for up to 8 hours per task.
Is GDPx the same as OpenAI's GDPval?
No. GDPval is OpenAI's 2025 benchmark of one shot deliverables across 44 occupations. GDPx is a separate SemiAnalysis benchmark that was designed partly as a response to GDPval, with interactive environments instead of prompts that contain all the context.
Which model has the highest GDPx score?
On the November 2, 2026 board, Claude Opus 5.5 leads at 12.0%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. If you leave out cybersecurity, biology and alignment, GPT-6 Astra and Opus 5.5 are roughly tied.
Why are GDPx scores so low?
The tasks are long, most are graded pass or fail against strict requirements, and finance and legal score almost zero under binary grading. SemiAnalysis designed the set so that even the strongest models solve only a small share.
Can I run GDPx myself?
Not fully. Each partner published one task with its rollouts, and the rest are private to prevent training on the test. You can read the public tasks and the full results on the GDPx site.
How often will GDPx be updated?
SemiAnalysis calls it a living benchmark and plans a new version every 3 to 6 months. Scores from different versions should not be compared directly.
