# SemiAnalysis GDPx explained: why the best AI agent scores 12 percent, what each point costs, and the caveats the headline leaves out > Source: > Published: 2026-11-05 17:08:51+00:00 # SemiAnalysis GDPx explained: why the best AI agent scores 12 percent, what each point costs, and the caveats the headline leaves out SemiAnalysis built GDPx, 206 long agent tasks across 11 expert fields, and the best model scored 12 percent. We worked out what each model pays per point, pulled the caveats SemiAnalysis states into the open, and walked through the one task that shows how strict the grading is. **This explains reporting by** [SemiAnalysis: GDPx, Measuring Frontier Model Capabilities Across the Most Valuable Domains (November 2, 2026), and the GDPx leaderboard](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities). Read the original first: [https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities) ## In one minute - GDPx has 206 tasks in 11 fields such as chip design, biology, finance, law and cybersecurity, with agents working up to 8 hours in realistic environments. - Claude Opus 5.5 leads at 12.0 percent and $51.41 a task. Gemini 4 Carbon scores 10.9 percent at $16.29. - Per point, the cheap models win: DeepSeek 4.1 Flash and GPT-6 Luna cost under 20 cents per point, Opus 5.5 about $4.28. - SemiAnalysis says safety classifiers pull down the scores of Opus 5.5 and GPT-6 Astra in cyber, biology and alignment. - The leaderboard text shows 13 of the 20 models charted, and SemiAnalysis plans a new version every 3 to 6 months. ## What GDPx measures SemiAnalysis published GDPx on November 2, 2026 as a benchmark for general agent work in high-value fields: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. The leaderboard lists 206 tasks from 14 data partners. Tasks run up to 8 hours by default. One finance task drops the agent into a deal room of about 2,000 files. Models run in their own harness where one exists, such as Claude Code or Codex, and in OpenCode otherwise. The headline score averages the 11 fields equally, and each partner equally within a field, using each task's own grading over 3 attempts. ## The leaderboard, with cost per point SemiAnalysis lists an average cost per task next to each score. Dividing one by the other gives what each point of GDPx costs. These are our numbers, computed from the leaderboard; SemiAnalysis does not publish this column. - DeepSeek 4.1 Flash 1017: 5.4 percent at $0.94 a task, $0.17 per point. - GPT-6 Luna: 5.1 percent at $0.92, $0.18 per point. - GPT 6.1 Sol: 8.4 percent at $9.80, $1.17 per point. - Gemini 4 Carbon: 10.9 percent at $16.29, $1.49 per point. - Gemini 4 Argon: 7.3 percent at $12.79, $1.75 per point. - Muse Wonder: 6.8 percent at $15.63, $2.30 per point. - Grok 4.7: 3.9 percent at $16.34, $4.19 per point. - Claude Opus 5.5: 12.0 percent at $51.41, $4.28 per point. - GLM 5.3: 2.7 percent at $11.79, $4.37 per point. - Claude Sonnet 5.5: 8.2 percent at $45.27, $5.52 per point. - GPT-6 Astra: 9.6 percent at $53.99, $5.62 per point. - Claude Fable 5.1: 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. Two readings follow. If you need the single best result, Opus 5.5 leads. If you run many attempts and keep the best, Gemini 4 Carbon gets 91 percent of Opus's 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 caveats, up front - Safety filters. 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. - Partial board. The leaderboard text marks itself partial and shows 13 of the 20 models in its cost chart. - Grading varies by field. 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. - Taste. SemiAnalysis says some fields rely on a single partner and that tasks were picked by QA and taste. - A moving target. GDPx is meant to change every 3 to 6 months, so today's scores will not compare cleanly to the next version. ## Where every model scored zero On binary grading, every model scored 0 percent in finance. 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 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. ## What one point demands: the quant task 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 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. That is the useful lesson for anyone running agents on real work. 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. ## What this means if you run agents - Long, expert, multi-hour jobs are still mostly out of reach, even for the best model. Use agents for parts of the job and keep a person on the final call. - Pick models by cost per useful result, not by top score alone. The spread between models is about 4 times on score and about 45 times on cost per point. - Treat security and biology numbers for filtered models as a floor, not a ceiling. ## Who is affected | Case | Status | |---|---| | Teams choosing a model for long agent tasks | Opus 5.5 leads on score; Gemini 4 Carbon and GPT 6.1 Sol lead on score per dollar among strong models. | | Security and biology workloads | Scores for Opus 5.5 and GPT-6 Astra are pulled down by safety classifiers, per SemiAnalysis. | | Finance and legal workloads | 0 percent binary pass rate for every model. | ## What to do - Compare models on cost per useful result for your own tasks. - Keep a person on final sign-off for multi-hour expert work. - Read the public GDPx tasks in your field before trusting a vendor's agent claims. ## What is still unknown - Per-field scores for each model. They are in charts that did not come through as text. - Exactly how cost per task is computed. The leaderboard says the cost chart uses output tokens; we did not confirm whether input tokens are included. - The 7 charted models missing from the leaderboard text. - How much the safety classifiers lower each model's score. SemiAnalysis states the effect but not its size. ## Sources - [SemiAnalysis: GDPx, Measuring Frontier Model Capabilities Across the Most Valuable Domains (November 2, 2026), and the GDPx leaderboard](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities) — the original report - [GDPx leaderboard](https://gdpx.semianalysis.com/) - [The Decoder: No AI agent scores above 12 percent on SemiAnalysis' new real-world work benchmark](https://the-decoder.com/no-ai-agent-scores-above-12-percent-on-semianalysis-new-real-world-work-benchmark/) [AI News Report](https://theainewsreport.com/)· every headline, every morning.