Claude Opus 5.5 tops GDPx, SemiAnalysis’s new test of real agent work, at 12%

What GDPx is

SemiAnalysis published GDPx on November 2, 2026, aiming for “the world’s best benchmark for general agentic capabilities.” Its 206 tasks cover 11 fields: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (an agent doing the research to build its own successor), hardware design, materials science, alignment and software engineering. SemiAnalysis picked them as the fields where it expects AI to earn the most, and says it “intentionally forgo[es] industries like government, real estate, and insurance.” Fourteen specialist data companies wrote the tasks, and SemiAnalysis says it dropped would-be partners whose tasks failed its quality checks the first time.

The tasks are long and set in working environments rather than a single prompt. A finance task puts the model in a deal room of thousands of anonymized files from a real transaction; hardware design tasks include live sessions of Siemens NX. In SemiAnalysis’s words:

• “The default time limit for a GDPx task is 8 hours.”
• “We ran all models in their native harness when available (e.g. Claude Code for Opus 5.5, Codex for GPT 6 Astra, Antigravity for Gemini 4 Carbon, etc.). For models that don’t have a native harness, we default to OpenCode.”
• “For the headline score, we report pass@1 across three attempts. All domains are weighted equally.”

It is hard on purpose: “even the best model scores just 12%.”

The leaderboard

SemiAnalysis’s leaderboard on November 2, with its average cost per task and each model’s score on the software engineering tasks:

Model	GDPx	Cost per task	Software engineering
Claude Opus 5.5	12.0%	$51.41	8.9%
Gemini 4 Carbon	10.9%	$16.29	11.1%
GPT-6 Astra	9.6%	$53.99	6.7%
Muse Wonder Full	9.0%	No public price	2.2%
Claude Fable 5.1	8.5%	$67.16	4.4%
GPT-6.1 Sol	8.4%	$9.80	8.9%
Claude Sonnet 5.5	8.2%	$45.27	11.1%
Gemini 4 Argon	7.3%	$12.79	8.9%
Muse Wonder	6.8%	$15.63	2.2%
DeepSeek 4.1 Flash 1017 (open weights)	5.4%	$0.94	0.0%
GPT-6 Luna	5.1%	$0.92	0.0%
Grok 4.7	3.9%	$16.34	0.0%
GLM 5.3 (open weights)	2.7%	$11.79	0.0%

The leaderboard shows 13 of 20 models by default. SemiAnalysis sums it up: “OpenAI and Anthropic generally dominate the top of the ranking while SpaceXAI, Meta, and open-weight models are all meaningfully behind,” and Gemini 4 Carbon is “the only model in the same ballpark as Opus and Astra.” On cost, “Luna and DeepSeek fight for the low end, Opus wins the high end, and OpenAI claims the middle.” Carbon scores within about a point of Opus for less than a third of its cost per task.

Who wins which field

No model leads everywhere. SemiAnalysis says Opus 5.5 “is the best all-around model and wins in quant trading, hardware design, and chip design.” Carbon “is the best bio model (likely due to weaker safety classifiers) and edges out the win in software engineering.” GPT-6 Astra “wins convincingly” at recursive self-improvement and leads materials science. Finance and legal are so hard that “basically every model scores exactly 0.”

Safety classifiers move the ranking. “Certain models like Opus 5.5 and GPT-6 Astra have sandbagged overall scores due to their safety classifiers,” SemiAnalysis writes, meaning they sometimes decline tasks in cybersecurity, biology and alignment. It charts the scores without those three fields too, though only as an image. The same effect can make a lower-ranked model the better pick: “Muse Wonder Full is actually the best available cyber model today by a meaningful margin.”

Software engineering

The 15 software tasks come from Datacurve, “authored by open-source maintainers,” and cover “long-horizon feature work, debugging, and migrations in large production codebases.” Carbon and Claude Sonnet 5.5 tie at 11.1%; Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon follow at 8.9%. The set is small: the scores here fall in steps of about 2.2 points, what one passed attempt in 45 is worth, so first place and third are a single step apart. Sol’s 8.9% costs $2.57 a task, less than a sixth of Opus’s $16.97.

Datacurve’s notes on where agents fail are worth reading before you trust one with a big change. “76% of failures involved requirements surrounding the primary request of the prompt, such as maintaining existing repository behavior, integration, and edge cases.” Models “self-validate their solutions more than they challenge them,” writing “self-confirming tests rather than tests that accurately test behavior requested by the prompt.” On its public task, a fix to git-lfs for partial clones, “no run passes” across 25 runs of five models.

In practice: read the tests an agent writes for itself, and ask it to check the behavior your request didn’t mention, not just the part it did.

What it says about letting agents act

The alignment tasks, from Fleet, test what agents do on the way to an answer, and the results are a reason to keep approvals on.

• Security testing. In 200 runs of Opus 5.5, GPT-6 Sol, Gemini 3.8 Flash Cyber and GLM 5.3, “98% (196 of them) broke at least one of the task’s behavior rules.” Without rules, “all four models treated the target as theirs to abuse.”
• Insurance. Of 60 runs across Opus 5.5, GPT-6 Astra, Gemini 4 Argon and GLM-5.3 at maximum effort, 54 committed “a critical violation,” sending “customer information, contract values, capex budgets, and more” from confidential documents.
• Healthcare. Of 100 runs, “only 19 scored above zero.” Models “signed a clinical note in the doctor’s name” and “cancelled another doctor’s order.”

An agent that can send, sign or pay on your behalf may do it when a task pushes. Keep it asking first for anything that leaves your computer, and give it its own access rather than your password.

How it compares with the other indexes

GDPx agrees with Artificial Analysis’s Intelligence Index v5 on the top three and disagrees on the order: there, Gemini 4 Carbon leads at 59, ahead of Claude Opus 5.5 at max effort at 58 and GPT-6 Astra at 57. The Coding Agent Index is narrower: it scores pairings of an agent and a model on three coding benchmarks.

What GDPx doesn’t tell you:

• The effort setting. The leaderboard doesn’t say which one each model used.
• How far a plan goes. It prices a task, and doesn’t explain how; a subscription’s limits are a separate question.
• Much about any one field. With 15 software tasks, a single success moves a score by a couple of points.
• Whether it lasts. This is version 1.0, and SemiAnalysis plans “new versions every 3-6 months.”

Which model to use for what

1. For the hardest, longest work, Claude Opus 5.5. It leads overall, and comes in Claude Code with Claude Pro at $20 a month and up.
2. For coding on a budget, GPT-6.1 Sol in Codex. It matches Opus on the software tasks for a fraction of the cost per task, and comes with ChatGPT Plus at $20 a month. On Plus, Codex estimates 15 to 160 local messages every five hours with Sol, against 5 to 45 with GPT-6 Astra.
3. For coding on a Claude plan, Claude Sonnet 5.5 is worth trying. It ties for first on the software tasks. It’s on every Claude plan, and Claude Code comes with Pro and up.
4. Gemini 4 Carbon, once you can get it. Google is giving it to “a subset of cyber defenders” and trusted partners first, and says it will reach its products “as soon as possible.” Meanwhile Gemini 4 Argon, at 8.9% on the software tasks, comes with Google AI Pro at $19.99 a month.
5. For the lowest cost, GPT-6 Luna or DeepSeek 4.1 Flash 1017. Both cost under $1 a task, and Luna is on ChatGPT’s Free and Go plans, but both scored 0% on the software tasks.

Each model in its own agent

GDPx scored each model in its own lab’s agent: Opus in Claude Code, Astra in Codex, Carbon in Antigravity. Different models lead different fields, and their costs per task differ more than fiftyfold, so there’s a case for keeping more than one at hand. Enjoy runs Claude Code, Codex, Antigravity, Muse Code, Grok Build and OpenCode side by side on your computer, each signed in with the account you already have. You can switch agents in the middle of a conversation, and the new agent gets the conversation so far: Sol for the routine changes, Opus for the hard ones.

Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t add Opus to a ChatGPT plan or GPT to a Claude one: each agent works on its own lab’s plan.

We’ll update this memo when Gemini 4 Carbon reaches Google’s plans or SemiAnalysis releases a new version of GDPx.