SemiAnalysis has published a new test for AI agents, and the headline result is humbling for the whole industry. The best model in the world completed just 12 percent of the work put in front of it. The more useful result, though, sits one row lower in the table, where a model that costs less than a third as much per task gets within 1.1 percentage points of the winner.
The benchmark is called GDPx, and SemiAnalysis released version 1.0 on November 2, 2026. It is built around 206 long tasks across 11 professional domains, assembled with 14 specialist data partners. The domains are chip design, hardware design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, materials science, alignment, and software engineering. Each task carries a default limit of eight hours, and models work in their native harness where one exists, Claude Code for Anthropic, Codex for OpenAI, and Antigravity for Google, with OpenCode as the fallback.
That design is the point. GDPx is not a quiz. It measures whether an agent can carry a piece of real professional work through to a graded result, and the headline score averages each model’s pass at first attempt result across three attempts, with every domain weighted equally. SemiAnalysis positions it as the successor to OpenAI’s GDPval, which it argues belongs to the pre-agent era, and it plans new versions every three to six months.
Claude Opus 5.5 leads, Gemini 4 Carbon leads on value
Anthropic’s Claude Opus 5.5 tops the debut leaderboard at 12.0 percent, at an average cost of 51.41 dollars per task and about 548,000 output tokens. Google’s Gemini 4 Carbon is second at 10.9 percent, at 16.29 dollars per task and about 594,000 output tokens. OpenAI’s GPT-6 Astra is third at 9.6 percent, at 53.99 dollars per task and about 275,000 output tokens.
| Model | GDPx 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 |
| Muse Wonder Full (Meta) | 9.0% | Not listed in the headline table |
| Claude Fable 5.1 (Anthropic) | 8.5% | $67.16 |
| Gemini 4 Argon (Google) | 7.3% | Not listed in the headline table |
| DeepSeek 4.1 Flash (DeepSeek) | 5.4% | $0.94 |
| GPT-6 Luna (OpenAI) | 5.1% | $0.92 |
| Grok 4.7 (xAI) | 3.9% | Not listed in the headline table |
| GLM 5.3 (open weight) | 2.7% | Not listed in the headline table |
Put those numbers on a per point basis and the story sharpens. Gemini 4 Carbon costs about 1.49 dollars per percentage point of GDPx score, against about 4.28 dollars for Claude Opus 5.5 and about 5.62 dollars for GPT-6 Astra. In plain terms, Carbon delivers about 91 percent of the leader’s score at about 32 percent of the leader’s cost per task. For any team running agents at volume, that gap matters far more than the 1.1 point gap at the top.
There is a catch, and it is a big one. Gemini 4 Carbon is not generally available yet. Google has limited it to a subset of cyber defenders so far, with phased access to additional trusted partners through its Fairwind program, as we reported when Google’s Gemini 4 Carbon topped the security charts. So the cheapest strong performer on this leaderboard is, for most buyers, still a model you cannot buy. That makes the table a planning document more than a shopping list, for now.
Domain winners tell a different story than the headline
The overall ranking hides sharp differences by field. SemiAnalysis reports Claude Opus 5.5 as the best all around model, winning quant trading, hardware design, and chip design. Gemini 4 Carbon is the best biology model, with a narrow win in software engineering. GPT-6 Astra wins recursive self-improvement and materials science. A buyer choosing a model for one kind of work should look at the domain score, not the headline, because the model in third place overall may be the best model for their specific job.
Two domains stand out for the opposite reason. Under strict binary grading, no model solved a single task in finance or legal, and SemiAnalysis describes the result across models in those fields as basically zero. That is a striking result for two fields full of structured, rules based work, and it is a reminder that these tasks were chosen to be hard. A low score here does not prove agents are useless at finance or legal work in general. It proves they cannot yet complete these particular long, graded tasks unsupervised.
Three caveats before you quote the leaderboard
First, GDPx is SemiAnalysis’s own test, run by SemiAnalysis, on tasks built with its partners. It has not been independently audited, and the score measures the harness as well as the model. A different harness or task set could move the ranking.
Second, SemiAnalysis itself cautions that the Opus 5.5 and GPT-6 Astra overall scores are held back by their safety classifiers in the cyber, biology, and alignment domains. In other words, part of what looks like a capability gap may be a refusal gap, where the model declines or filters work its safety systems flag.
Third, cost per task depends on how much a model writes. Gemini 4 Carbon used about 594,000 output tokens per task, more than Opus 5.5 at about 548,000, yet still cost far less, which tells you its per token price is doing the heavy lifting. OpenAI’s GPT-6 Astra used far fewer tokens, about 275,000, but still cost 53.99 dollars per task. Token counts alone do not predict the bill.
What this means for AI buyers
The practical takeaway is to stop shopping on headline benchmarks alone. If you are deploying agents on real work, ask three questions that GDPx makes concrete. What does the model score in your domain, not overall? What does a completed task cost, not a million tokens? And can you actually get access to the model at that price, or is it gated behind a partner program?
On those questions, the November 2 leaderboard gives a clear interim answer. Claude Opus 5.5 is the strongest general agent in this test, Gemini 4 Carbon is the value leader among the top three, and GPT-6 Astra is close behind on score but expensive per task. The cheapest models in the table, DeepSeek 4.1 Flash at 0.94 dollars per task and GPT-6 Luna at 0.92 dollars, score only 5.4 percent and 5.1 percent, so cheap is not automatically good value either. The sweet spot, when Carbon opens up more widely, will be the model that holds its domain score while keeping its per task cost near Carbon’s level.
That pricing logic also applies to earlier launches. We recently compared OpenAI’s GPT-6.1 Terra at half the price of its predecessor, and open source agents, like DeepSeek’s newly open sourced coding agent, will keep pushing the market to judge models in dollars per finished task, not demo videos.
SemiAnalysis publishes the full leaderboard at gdpx.semianalysis.com, and the next version is due within three to six months. If Carbon’s cost advantage survives a wider release, and if its domain wins hold up on a fresh task set, the next leaderboard may have a different winner. For now, the honest summary is this: the best agent scores 12 percent, the best value agent scores 10.9 percent, and the distance between them is 35 dollars per task.
