---
type: "post"
id: "semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569"
slug: "semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569"
title: "SemiAnalysis launches GDPx, a 206-task agentic benchmark across 11 high-value domains; top model scores 12%"
description: "SemiAnalysis introduced GDPx, a benchmark of 206 long-horizon tasks across 11 domains intended to measure frontier models on economically valuable work."
retrieval_nugget: "SemiAnalysis introduced GDPx, a benchmark of 206 long-horizon tasks across 11 domains intended to measure frontier models on economically valuable work. Its leading model scored 12%, a result that underscores how far current systems remain from reliable performance on difficult professional tasks—and why benchmark design and scoring conditions matter."
published_at: "2026-11-03"
updated_at: "2026-11-07"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["claude","gemini","claude-code"]
entities: ["SemiAnalysis","GDPx","Claude Opus 5.5","Gemini 4 Carbon","GPT-6 Astra","DeepSeek 4.1 Flash","Claude Code","Codex"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "same_event"
prior_refs: ["https://newruntime.com/posts/semianalysis-releases-gdpx-an-agentic-benchmark-across-11-56f9ba/"]
source_url: "https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities"
source_urls: ["https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569/","markdown":"https://newruntime.com/posts/semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569.md","json":"https://newruntime.com/posts/semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569.json"}
markdown_url: "https://newruntime.com/posts/semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569.md"
---

# SemiAnalysis launches GDPx, a 206-task agentic benchmark across 11 high-value domains; top model scores 12%

## Retrieval answer

SemiAnalysis introduced GDPx, a benchmark of 206 long-horizon tasks across 11 domains intended to measure frontier models on economically valuable work. Its leading model scored 12%, a result that underscores how far current systems remain from reliable performance on difficult professional tasks—and why benchmark design and scoring conditions matter.

## Summary
SemiAnalysis introduced GDPx, a benchmark of 206 long-horizon tasks across 11 domains intended to measure frontier models on economically valuable work. Its leading model scored 12%, a result that underscores how far current systems remain from reliable performance on difficult professional tasks—and why benchmark design and scoring conditions matter.

## What happened
SemiAnalysis describes GDPx as a benchmark for measuring general agentic capabilities in domains where it believes frontier intelligence could have high economic value. The 11 domains are chip design, cybersecurity, biology, finance, legal work, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. The authors intentionally exclude areas such as government, real estate, and insurance, arguing that a general benchmark based on broad economic or occupational categories would miss the sectors where frontier models may have the highest impact.

The supplied launch material reports 206 tasks and collaboration with 14 specialized data partners. Tasks are designed to run in realistic environments, with examples including a private-equity deal room containing about 4,000 files, live multi-service cyber ranges, and SolidWorks or NX sessions. The default task time limit is eight hours. Models are evaluated in their native harnesses where possible, including Claude Code, Codex, and Antigravity; other models use OpenCode. Scores average three attempts, with domains weighted equally.

On the reported overall results, Claude Opus 5.5 leads at 12.0%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. The best open-weight model listed is DeepSeek 4.1 Flash at 5.4%. SemiAnalysis says finance and legal tasks are so challenging that essentially every model scores zero there under binary grading. The article also reports domain-specific strengths: Opus 5.5 leads in quant trading, hardware design, and chip design; Gemini 4 Carbon leads in biology and narrowly wins software engineering; GPT-6 Astra leads in recursive self-improvement and materials science. The authors say safety classifiers may have depressed some models’ aggregate scores in cybersecurity, biology, and alignment, and show a separate ranking excluding those domains.

GDPx is accessible through its official site, linked by SemiAnalysis in the article. The supplied material does not specify a price or broader access conditions. The article says SemiAnalysis plans new versions every three to six months and reports that its quality-assurance process flagged nearly half of Terminal Bench 4.0 tasks as broken, a comparison offered as context for the difficulty of maintaining reliable evaluation tasks.

## What the archive adds
This is a new fact on a known story, not a wholly new benchmark thread. Our earlier [GDPx Field Note](https://newruntime.com/posts/semianalysis-releases-gdpx-an-agentic-benchmark-across-11-56f9ba/) covered the benchmark’s launch, scope, and 12% leading result. The supplied material adds more detail on its task environments, execution setup, per-domain outcomes, and the role of safety classifiers.

## Trend signal
GDPx supports the pattern that agent evaluation is moving toward completed work rather than isolated model calls. Its long-horizon tasks run in work-like environments and are judged by task outcomes, though the low scores and task-quality caveat show how immature that measurement remains. See [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/).

## Why now
SemiAnalysis argues that widely reported general-purpose knowledge-work benchmarks were designed before contemporary agents were common and therefore do not adequately reflect agent task design. GDPx’s launch attempts to fill that gap with environments and tasks built around extended agent work.

## Connections to other events
The result sits alongside our earlier coverage of [GDPval-AA and other agentic benchmarks](https://newruntime.com/posts/new-benchmarks-cognition-s-frontiercode-2-0-merge-a78e19/), including Terminal-Bench-Science. Those efforts likewise assess agents on tasks that produce work artifacts or require extended execution, but GDPx emphasizes a broader set of high-value technical and professional domains.

## Why it matters
For teams deciding whether to delegate work to an agent, a low aggregate score is a reminder not to infer reliable completion from a convincing demonstration or a strong result on a narrow coding test. GDPx also suggests that capability is uneven: one model may be stronger in chip design or quant trading while another leads in software engineering or biology. That makes workload-specific evaluation more useful than choosing a system from a single leaderboard position.

The benchmark is not a direct forecast of workplace productivity. Its 12% leader result combines task selection, environment setup, scoring rules, and model behavior. Still, the scale and duration of the tasks make it a useful prompt for teams to identify their own consequential workflows, define observable completion criteria, and test agents against representative cases before assigning production work. Earlier coverage: [SemiAnalysis releases GDPx, an agentic benchmark across 11 high-value domains where the best model scores 12%](https://newruntime.com/posts/semianalysis-releases-gdpx-an-agentic-benchmark-across-11-56f9ba/).

## Limits and uncertainty
The reported figures do not establish how models perform on all real-world work, nor do they show that GDPx’s domain weighting matches any particular organization’s priorities. A zero under binary grading can conceal partial progress, while a score can depend on environment details and task construction. SemiAnalysis itself raises safety classifiers as a possible cause of lower results in three domains. The supplied material also does not give full per-task results, confidence intervals, or enough detail to independently reproduce the runs. Its reported quality-assurance concern about another benchmark illustrates a broader measurement risk, not proof that GDPx tasks are defective.

## Skill to build now
Choose one consequential workflow your team might delegate to an agent and turn it into a small evaluation set: representative inputs, a realistic working environment, a clear completion test, and a record of failures across repeated runs. GDPx’s design highlights why tool access and long task horizons should be part of the test—not just the final answer. Keep partial credit separate from verified completion so a polished attempt does not get mistaken for a successful task.

## Reader takeaway
GDPx’s headline is not simply that one model ranks first; it is that the leader scores only 12% on a demanding set of long-horizon tasks. Treat that as evidence of a large reliability gap, while reading the number in context: domain mix, binary grading, task quality, and safety behavior all influence what the score means.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- topic hub: [Claude Code](https://newruntime.com/topics/claude-code.md): Explore the claude-code topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)