GDPx Benchmark: Opus 5.5 vs Gemini 4 Carbon

SemiAnalysis released GDPx v1.0, a benchmark of 206 long-running agentic tasks across 11 domains built with 14 partners, and found that the best model, Anthropic's Claude Opus 5.5, passed only 12.0% of tasks at $51.41 per task. Google's Gemini 4 Carbon scored 10.9% at $16.29 per task and led biology at 25.6%, though SemiAnalysis attributed that lead partly to weaker safety classifiers and noted that Opus 5.5 and GPT-6 Astra appeared to sandbag on cyber, biology and alignment tasks. No model passed a single finance or legal task under binary grading, with the best finance run passing 8 of 11 required checks across 25 attempts.

On 2 November 2026, SemiAnalysis released GDPx v1.0, a benchmark of 206 agentic tasks across 11 domains, and the best model, Claude Opus 5.5, scored 12.0% GDPx leaderboard https://gdpx.semianalysis.com/ . Our verdict: Opus 5.5 is the better general-purpose pick, while Gemini 4 Carbon 10.9% overall at $16.29 per task, against $51.41 for Opus 5.5, per the official leaderboard is the stronger and cheaper choice for biology workloads, with a caveat about why it wins there. The headline is not that one model beats another. It is that the strongest agents on the market still fail close to nine in ten professional-grade tasks when the work is hard, private and graded without sympathy. Last verified: 3 November 2026. Figures come from the GDPx site https://gdpx.semianalysis.com/ , which showed data updated on 1 to 2 November. The GDPx benchmark is a set of long-running agent tasks built with 14 partners that each run a specialised reinforcement-learning environment. Each model gets a workspace, an eight-hour limit and three attempts. Scores are pass@1 averaged, with all 11 domains weighted equally and each partner weighted equally within a domain methodology https://gdpx.semianalysis.com/ . The domains span chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis positions it as a successor to OpenAI's GDPval, which used 1,320 tasks across 44 occupations TechTimes https://www.techtimes.com/articles/329945/20261103/gdpx-benchmark-best-frontier-ai-scores-12-real-work-zero-finance-legal.htm . Its argument is that occupational-taxonomy benchmarks aim at the wrong target, because early AI revenue will concentrate where frontier intelligence pays back fastest. On the published leaderboard, the spread between first and third is small and the cost spread is large leaderboard https://gdpx.semianalysis.com/ : | Model | Score | Cost per task | |---|---|---| | Claude Opus 5.5 Anthropic | 12.0% | $51.41 | | Gemini 4 Carbon Google | 10.9% | $16.29 | | GPT-6 Astra OpenAI | 9.6% | $53.99 | | Claude Fable 5.1 | 8.5% | $67.16 | | Muse Wonder Full Meta | 9.0% | not public | | GPT-6.1 Sol | 8.4% | $9.80 | | DeepSeek 4.1 Flash 1017 open weights | 5.4% | $0.94 | Gemini 4 Carbon delivers about 91% of the leader's score at roughly a third of the price, a figure derived from the two published numbers above. Claude Opus 5.5 ran in Claude Code, GPT-6 Astra in Codex and Gemini 4 Carbon in Antigravity, each in its native harness, which is worth remembering before comparing scores to your own setup. For other head-to-head views, see our GPT-6 Astra vs Claude Fable 5.1 verdict https://dev.to/articles/gpt-6-astra-vs-claude-fable-5-1-verdict-2026 and the OS 2.0 agent benchmark https://dev.to/articles/os-2-0-agent-benchmark . Under binary grading, a task passes only when every required check passes, and no model managed that in finance or legal SemiAnalysis https://gdpx.semianalysis.com/ . In the finance task "Bellwood-QoE", the best single run passed 8 of 11 required checks across 25 attempts. In the public legal task, a Series B redline, none of the 20 rollouts passed on a binary basis partial credit averaged 18% GDPx https://gdpx.semianalysis.com/ . The practical reading: a single missed clause or reconciliation still fails the work, so a reviewer remains necessary. On the published numbers, yes: Gemini 4 Carbon scores 25.6% in biology, the highest of any model tested GDPx https://gdpx.semianalysis.com/ . The caveat matters. SemiAnalysis found that Opus 5.5 and GPT-6 Astra appeared to sandbag on cyber, biology and alignment tasks, and attributes Gemini's lead to weaker safety classifiers, stating this is its own interpretation, not a Google statement. Its secondary leaderboard, which excludes cyber, biology and alignment, changes the order: GPT-6 Astra reaches 11.8%, Opus 5.5 11.7% and Gemini 4 Carbon drops to 9.2% GDPx https://gdpx.semianalysis.com/ , data updated 2 November; corroborated by roboticaguide.com https://roboticaguide.com/gdpx-benchmark-results-opus-5-5-gemini-4-carbon/ . So the biology lead is partly a question of how willing a model is to run the experiment, not only how well it reasons. LatchBio built the 13 biology tasks, 12 private and one public, covering single-cell and spatial transcriptomics, genomics and metagenomics, drug discovery, toxicology and biosecurity. The data is real: tumour biopsies, donor pancreatic islets and public cell atlases modified to prevent memorisation, with each task written by a scientist, checked by two others and graded deterministically BioBench https://benchmarks.bio/ . LatchBio documented three recurring failure modes: In the public antibody task, models could not reconcile a chip-based binding assay with a cell potency assay, and Opus 5.5 hit that failure mode every time. If you deploy agents in a lab, build these checks into review. Our pieces on multi-agent science workbenches https://dev.to/articles/claude-science-ai-workbench-multi-agent-orchestrators and agentic AI in health care https://dev.to/articles/agentic-ai-in-health-care-compared-2026 cover the surrounding tooling. Choose Claude Opus 5.5 if you need one model across varied professional tasks. It leads overall and was the best all-around performer, scoring 27.4% in quant trading and 18.7% in hardware design per-field scores via GDPx https://gdpx.semianalysis.com/ ; corroborated by roboticaguide.com https://roboticaguide.com/gdpx-benchmark-results-opus-5-5-gemini-4-carbon/ . Choose Gemini 4 Carbon if the work is biology or software engineering 11.1%, per the GDPx https://gdpx.semianalysis.com/ field scores; corroborated by roboticaguide.com https://roboticaguide.com/gdpx-benchmark-results-opus-5-5-gemini-4-carbon/ and cost matters, and accept that part of its edge may reflect looser guardrails. Choose GPT-6 Astra for recursive self-improvement 24.9% and materials science 21.6% , per the GDPx https://gdpx.semianalysis.com/ domain table corroborated by roboticaguide.com https://roboticaguide.com/gdpx-benchmark-results-opus-5-5-gemini-4-carbon/ . For pure coding, see our real SWE benchmark comparison https://dev.to/articles/which-llm-is-best-for-coding-real-swe-benchmark-compared . Limits to keep in mind: most tasks are private, so results cannot be independently reproduced, and equal domain weighting means a niche domain moves the average as much as a large one. Q: What is the GDPx benchmark? A: GDPx is a SemiAnalysis benchmark of 206 agentic tasks across 11 domains, built with 14 environment partners and released on 2 November 2026. Q: Which model scored highest on GDPx? A: Claude Opus 5.5 from Anthropic, at 12.0% and $51.41 per task, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6% leaderboard https://gdpx.semianalysis.com/ . Q: Why is a 12% top score meaningful? A: Tasks are long, private and graded on deterministic checks, so a pass means the work was fully correct. It shows that unsupervised expert-level work is still out of reach for current agents. Q: Why did Gemini 4 Carbon win in biology? A: It scored 25.6% GDPx https://gdpx.semianalysis.com/ , and SemiAnalysis suggests weaker safety classifiers contributed, while noting this is its own reading and not a statement from Google. Q: Is GDPx a replacement for GDPval? A: SemiAnalysis pitches it as the successor, arguing that occupation-based benchmarks miss where early AI revenue will concentrate. Whether the community adopts it is not yet known. This article was produced with AI assistance and checked against the primary sources linked above. Corrections will be logged here if figures change.