What happened
GDPx presents itself as a benchmark of frontier model capabilities across high-value domains. Its version 1.0 covers 206 tasks in 11 areas: chip design, cybersecurity, biology, finance, legal work, quantitative trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. SemiAnalysis says the benchmark was built with 14 specialized data partners, which supplied domain-specific task environments.
The leaderboard gives each domain equal weight. Scores aggregate an average of three attempts, with tasks graded using binary or continuous scoring. This structure is intended to represent work in which an agent must act inside an environment rather than simply answer a question. The supplied benchmark material does not provide enough detail to reconstruct every task or scoring rubric, but it identifies the domains and reports overall results for participating models. The site also provides views for comparing models by domain and a cost-versus-performance chart.
Claude Opus 5.5 leads the v1.0 table at 12.0%. Gemini 4 Carbon scores 10.9%, GPT-6 Astra 9.6%, Muse Wonder Full 9.0%, Claude Fable 5.1 8.5%, GPT 6.1 Sol 8.4%, and Claude Sonnet 5.5 8.2%. Among the open models listed, DeepSeek 4.1 Flash scores 5.4% and GLM 5.3 scores 2.7%. These results are low in absolute terms: even the leading model does not complete most of the benchmark's scored work successfully. That makes the leaderboard a measure of current capability on this task set, rather than evidence of broadly solved autonomous work.
GDPx also displays average cost per task, letting readers compare performance with price. The supplied launch material gives Claude Opus 5.5 at $51.41 per task and Gemini 4 Carbon at $16.29; DeepSeek 4.1 Flash is listed at $0.94 per task. One model, Muse Wonder Full, is marked as having no public price. Each partner open-sourced rollouts for one task, while the remaining rollouts are private to reduce benchmark hill-climbing. The benchmark site exposes domain and model views, but the supplied material does not specify a broader access or pricing policy for using GDPx itself.
What the archive adds
This is a new benchmark release, but it extends an existing thread on evaluating agentic work. New benchmarks: Cognition's FrontierCode 2.0 and Terminal-Bench-Science covered task-based tests for software engineering and scientific agents. GDPx adds a wider set of domains and task-cost comparisons; it is a new fact in the broader benchmark story, not a restatement of those releases.
Trend signal
GDPx supports the pattern that agent economics should be measured as cost per verified completed task, not just cost per token or model call. Its leaderboard places task-level cost alongside scores, showing a meaningful price-performance comparison: Gemini 4 Carbon scores close to the leader at a reported lower task cost. Agent economics moves to completed work is therefore directly relevant, although GDPx does not by itself establish how these benchmark costs translate to production deployments.
Why it matters
For teams considering domain-specific agents, a single general-purpose benchmark score can hide large differences in task type and operating cost. GDPx's domain breakdown gives evaluators a reason to ask whether a model's strengths match their workflow, while its task-cost figures make the trade-off more concrete than a model-call price alone. A model that is near the top on the aggregate table may still be the wrong candidate for a particular domain, and a cheaper candidate may deserve testing when its score is sufficiently close.
The caution is as important as the comparison. A 12.0% top score signals that current models fail most benchmark work under GDPx's evaluation, not that an agent is ready to own a full professional workflow. The three-attempt average and limited public rollouts leave open questions about variance, reproducibility, and how much of the result can be independently inspected. Teams should treat the leaderboard as a screening signal, then run representative evaluations on their own tasks with explicit success criteria and human review. Later coverage: SemiAnalysis launches GDPx, a 206-task agentic benchmark across 11 high-value domains; top model scores 12%.
What happens next
- GDPx will add a useful reference point in model-selection discussions for specialist agents. (3 months, 70%). Watch for: newly reported model entries or domain-level comparisons.
- Teams will need local evaluations before treating a high GDPx placement as a deployment decision. (6 months, 80%). Watch for: published replication results or task-specific comparisons.
Limits and uncertainty
The supplied material does not document the full task set, each grading rubric, model configuration, or the distribution of scores across the three attempts. Only one rollout per partner is open-sourced; the rest remain private to discourage tuning against the benchmark. The material also does not establish that benchmark task costs will match costs in production, where context, retries, tools, and infrastructure can differ. Equal weighting is a benchmark design choice, not proof that every domain has equal relevance to a given organization. Scores should not be read as a direct measure of general job performance or as evidence that any model is reliable enough for unsupervised professional work.
Skill to build now
Create a small evaluation sheet for one recurring agent workflow this week. Define the task input, the observable completion condition, a quality check, the allowed retry count, and the fully loaded cost of each run. Test at least two model options on the same cases. GDPx is a reminder that model price and model capability need to be compared at the completed-task level, not inferred from token pricing or a general leaderboard alone.
Reader takeaway
GDPx is a useful new comparison because it puts agent performance across 11 domains beside cost per task. Its headline is not that agents have mastered these fields: the top score is 12.0%. Use the leaderboard to identify candidates and questions for local evaluation, not as a substitute for testing the work your team actually needs done.