GDPx agent benchmark: cost per point for 12 models, and the safety-filter

SemiAnalysis published GDPx on November 2, 2026 under the title GDPx, Measuring Frontier Model Capabilities Across the Most Valuable Domains. It holds 206 tasks from 14 data partners in 11 fields, and a task can run up to 8 hours. Claude Opus 5.5 leads at 12.0 percent and $51.41 a task. theainewsreport.com's report on November 5, 2026 divides cost by score to price each point, a column SemiAnalysis does not publish.

This piece is for teams choosing a model for long agent work in chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. It covers equal-field scoring, cost per point for 12 priced models, the zero-score fields, one public quant grader, and the safety-filter limits.

GDPx agent benchmark: what actually changed

Where a model has its own harness, the run uses that harness, such as Claude Code or Codex, and OpenCode otherwise. One finance task drops the agent into a deal room of about 2,000 files. The Decoder says no AI agent scores above 12 percent on SemiAnalysis' new real-world work benchmark. The leaderboard text shows 13 of the 20 models in its cost chart. The headline score averages the 11 fields equally, and each partner equally within a field, using each task's own grading over 3 attempts.

DeepSeek 4.1 Flash 1017 scores 5.4 percent at $0.94 a task, $0.17 per point. GPT-6 Luna scores 5.1 percent at $0.92, $0.18 per point. Both are under 20 cents per point. GPT 6.1 Sol scores 8.4 percent at $9.80, $1.17 per point. Gemini 4 Carbon scores 10.9 percent at $16.29, $1.49 per point. Gemini 4 Argon scores 7.3 percent at $12.79, $1.75 per point. Muse Wonder scores 6.8 percent at $15.63, $2.30 per point. SemiAnalysis does not publish this per-point column.

GDPx agent benchmark: how it works

Finance and legal use binary pass or fail, with partial scores also reported. Biology is deterministic. Hardware mixes verifiers with an LLM judge against expert rubrics. SemiAnalysis says some fields rely on a single partner and that tasks were picked by QA and taste. Agents can work for up to 8 hours by default in realistic environments. Per-field scores for each model are in charts that did not come through as text, leaving the equal-weighted headline as the written grade.

The public quant task aapl-target-winsorization asks the agent to beat a reference trading strategy on hidden days. A strategy scores 1 only if a one-sided paired t-test on daily profit gives p below 0.01, and 0 if p is above 0.5. The reference is scaled up when the agent takes more risk. 10 of 12 runs made more than the reference's $1,399 out of sample, yet only 2 reached a score of 0.5. One Claude Opus 5.5 run made $2,540 at 1.3 times the risk and scored 0.23. A Grok 4.7 run scaled its predictions by 1.15, took 1.46 times the risk, and scored 0.

GDPx agent benchmark: why it matters

Grok 4.7 scores 3.9 percent at $16.34 a task, $4.19 per point. Claude Opus 5.5, at 12.0 percent and $51.41, comes to about $4.28 per point. GLM 5.3 scores 2.7 percent at $11.79, $4.37 per point. Claude Sonnet 5.5 scores 8.2 percent at $45.27, $5.52 per point. GPT-6 Astra scores 9.6 percent at $53.99, $5.62 per point. Claude Fable 5.1 scores 8.5 percent at $67.16, $7.90 per point. Muse Wonder Full scores 9.0 percent but has no public price, so it has no cost per point.

For the single best result, Claude Opus 5.5 leads. If you run many attempts and keep the best, Gemini 4 Carbon gets 91 percent of that score for about a third of the cost per task, and GPT 6.1 Sol gets 70 percent of it for under a fifth. The spread between models is about 4 times on score and about 45 times on cost per point. Long, expert, multi-hour jobs are still mostly out of reach. Use an agent for parts of the job and keep a person on the final call.

GDPx agent benchmark: who is affected

Teams choosing a model for long agent tasks have two rankings. Claude Opus 5.5 leads on score. Gemini 4 Carbon and GPT 6.1 Sol lead on score per dollar among strong models. SemiAnalysis writes that models like Opus 5.5 and GPT-6 Astra have sandbagged overall scores due to their safety classifiers, which matters in the cyber, biology, and alignment fields. It describes Muse Wonder Full as the best cyber model once classifiers are considered. Treat security and biology numbers for filtered models as a floor, not a ceiling.

Finance and legal workloads have a 0 percent binary pass rate for every model. One finance task, Bellwood-QoE, went unsolved across 25 attempts, and the best run passed 8 of 11 checks. In legal, 20 rollouts produced no usable contract redline, though partial scores averaged 44 percent for Claude Opus 5.5 and 25 percent for GPT-6 Astra. In hardware design, none of 12 rollouts reached a verified full pass. In the recursive self-improvement task, none of 20 rollouts reliably found the cause of a broken RL harness.

GDPx agent benchmark: what to watch

SemiAnalysis plans a new GDPx version every 3 to 6 months, so today's scores will not compare cleanly to the next version. The leaderboard text shows 13 of the 20 models, leaving 7 charted models missing from the published leaderboard text. Per-field scores for each model are in charts that did not come through as text. Some fields rely on a single partner. Tasks were picked by QA and taste, and GDPx is meant to change every 3 to 6 months.

The leaderboard says the cost chart uses output tokens, and it is not confirmed whether input tokens are included. How much the safety classifiers lower each model's score is unstated, even though SemiAnalysis states the effect. Beating a number is not the same as beating it reliably at the same risk, and a strict grader can tell the difference when a demo cannot. Compare models on cost per useful result for your own tasks, and keep a person on final sign-off for multi-hour expert work.