Best AI Agent Scores Just 12% on GDPx in 2026

The best AI agent in the world just scored 12% on real expert work. That is the headline from SemiAnalysis’s GDPx benchmark, published November 2, 2026: Anthropic’s Claude Opus 5.5 finished first with a 12.0% score, ahead of Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%. No model cleared 13%.

If that number shocks you, good — it should change how you hand work to AI agents. But it should not convince you that agents are useless. GDPx measures something very specific: long, messy, professional jobs that take up to eight hours each, graded pass or fail. It is not a test of writing a decent email or summarizing a document. The useful question is not “is 12% bad?” but “what kind of work does that 12% cover, what does each success cost, and where should you still keep a human in charge?” This article answers all three, with the leaderboard, the cost math, and a practical framework for deciding what to delegate.

What GDPx actually tests

GDPx v1.0 is a benchmark built by SemiAnalysis researchers Max Kan, Andrew Megalaa, and Dylan Patel. It contains 206 tasks across 11 domains chosen for economic value rather than for covering the whole economy: chip design, hardware design, cybersecurity, biology, finance, legal, quantitative trading, recursive self-improvement, materials science, alignment, and software engineering. Fourteen specialist data partners helped build the tasks.

Three design choices make GDPx different from the benchmarks you usually see on model launch pages:

• Agents work, they don’t just answer. Each model runs inside a simulated workplace in its native agent setup — Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon — and has up to eight hours per task. GDPx replaces GDPval, OpenAI’s September 2025 benchmark, which handed models all their context in the prompt and scored them without interaction. GDPx tasks force the agent to go find things, use tools, and carry a job through to a graded result.
• Grading is mostly pass or fail. A plausible-looking draft that fails the acceptance test does not pass. The headline score is pass@1 averaged over three attempts, with all 11 domains weighted equally.
• The tasks are hard on purpose. One finance task recreates a private equity deal in which a fund buys roughly 62% of a snack manufacturer for $880 million plus an earnout, spread across about 4,000 anonymized deal files the agent has to work through. That is the kind of job where a confident-sounding wrong answer is worse than no answer at all.

Two caveats matter before you quote the leaderboard anywhere. First, these are SemiAnalysis’s own measurements, run by the firm that built the benchmark; no outside group has independently reproduced them yet, and many tasks are private so they cannot leak into training data. Second, SemiAnalysis plans to refresh GDPx every three to six months, and seven of the 20 models tested have not had their scores published yet. Treat v1.0 as a first snapshot, not a final verdict.

The GDPx leaderboard: first place is 12%

Here is the top of the published leaderboard, with 13 of the 20 tested models public:

The gap at the top is small: only 1.1 points separate Opus 5.5 from Gemini 4 Carbon, and 2.4 points separate it from GPT-6 Astra. The gap between the leaders and the rest is large. The best open-weight model on the board, GLM 5.3, scored 2.7% — less than a quarter of the leader’s score. If you are picking a model for hard agent work, the leaderboard says the frontier is a small club right now, and the open-weight options are not in it yet.

One more number deserves attention: SemiAnalysis reports that every model scored zero on finance and legal tasks under binary grading. Those two domains only show nonzero numbers because SemiAnalysis switched to partial credit there. Keep that in mind when you read the domain breakdown below — the finance and legal figures are not comparable with the rest on equal terms.

The number that matters more: cost per success

A benchmark score tells you how often an agent finishes the job. Your budget cares about something else: how much you spend per finished job. SemiAnalysis publishes an average cost per task attempt, so the cost-per-success math is simple division — attempt cost divided by the pass rate. Here is what it gives for the three models with published attempt costs:

Read that table twice. Opus 5.5 wins on score and still costs about $428 for every task it actually completes, because you pay for the eight failed attempts too. Gemini 4 Carbon lands within 1.1 points of the leader at roughly a third of the cost per attempt — about $149 per success. That is the value story hiding inside the headline, and it is the reason the leaderboard alone is a bad buying guide.

There is a catch, and it is a big one: Gemini 4 Carbon is not generally available yet. Google has limited it so far to a subset of cyber defenders, with phased access to additional trusted partners through its Fairwind program, and no general availability date has been announced. So the cheapest path to a good GDPx score is one most teams cannot buy today. If you are budgeting an agent pipeline this month, plan around the models you can actually use — Opus 5.5, GPT-6 Astra, and their smaller siblings — and treat Carbon’s $149 as a preview of where prices may head, not a price you can pay.

For context on how agent costs add up in real freelance and small-business work, see our breakdown of AI agent costs for US freelancers.

Which model wins where: the domain routing table

The overall leaderboard hides the most useful part of GDPx: different models win different kinds of work. If you route jobs by domain instead of defaulting to one model for everything, you get more out of the same budget. Here is the domain picture SemiAnalysis published:

Two practical lessons fall out of that table. First, “best model” is the wrong question; “best model for this job” is the right one. A team doing chip-design support and biology literature review has no reason to run both jobs on the same model. Second, the cyber result is a warning against reading any single number in isolation: Muse Wonder Full’s cyber lead comes with the caveat below about safety filters cutting the other way for its rivals.

The safety-filter caveat you should not skip

SemiAnalysis states plainly that safety classifiers drag down the scores of Opus 5.5 and GPT-6 Astra in cybersecurity, biology, and alignment — the models refuse or cut short tasks their own filters flag. That means the headline scores measure two things at once: what the model can do, and what its gatekeeper lets it do.

One analysis of the published charts reports that excluding those three domains entirely would flip the top of the table — Astra would edge ahead of Opus 5.5 by about a tenth of a point, and Carbon would fall from second to fifth. Treat that as reported analysis, not as SemiAnalysis’s own headline, but the direction is what matters: if your work sits in cyber, biology, or alignment-adjacent areas, the published ranking may understate what these models can do once their safety settings are configured for your legitimate use case, and it may overstate the gap between models. Check the domain scores for your field, not just the overall number.

What 12% does not mean

A 12% score on deliberately brutal expert tasks is easy to misread in both directions. Here is what it is not:

• It is not “AI can only do 12% of the economy.” GDPx domains were picked for high return on intelligence, not for share of GDP, and the tasks are the hard tail of professional work. Routine, well-defined jobs — drafting, extraction, summarization, first-pass research — score far higher on other evaluations and in everyday use. GDPx measures the ceiling, not the average.
• It is not “agents are useless.” Completing one in eight eight-hour expert tasks unsupervised is, honestly, further than most people expected agents to be by late 2026. The right reaction is narrower delegation, not no delegation.
• It is not a pure model ranking. GDPx scores the model plus its harness plus its safety settings, on tasks built with commercial partners. A different harness or a different filter configuration could move a score. That is another reason to test models on your own tasks before committing budget — your harness is the one that counts.
• It is not permanent. GDPx refreshes every three to six months. The first version is a baseline; the trend across versions will tell you more than any single score.

A task-size framework: what to hand an agent, and what to keep

Here is the original-value part of this article — a simple framework built from what GDPx actually shows. Sort every job you are tempted to delegate into three tiers, and treat each tier differently:

Tier 1: Short, well-defined tasks — delegate freely

One clear deliverable, one source of truth, easy to check. Examples: summarize this report, draft this email from these notes, extract these figures into a table, write a first draft from this outline. Agents handle these well, failures are cheap and visible, and a quick human read catches them. This is where most small businesses should be getting value today.

Tier 2: Multi-step jobs with checkpoints — delegate with guardrails

Several steps, some judgment calls, real consequences if a middle step goes wrong. Examples: research a vendor shortlist, prepare a client proposal from scattered files, reconcile a month of transactions. Run the agent in stages, check each stage before the next starts, and never let it send, file, pay, or publish without a human approving the final output. Our guide to agent guardrails for small businesses walks through the no-stop-condition problem — an agent that keeps going when it should have asked — which is exactly how Tier 2 jobs go sideways.

Tier 3: Eight-hour expert jobs — do not hand over the whole job

This is GDPx territory: a full due-diligence review, a chip-design task, a legal analysis across thousands of files. The data says the best agent fails about seven times in eight here, at $50-plus per attempt. Break the job into Tier 1 and Tier 2 pieces instead — one agent pass to organize the files, a human to set the analysis plan, an agent to draft one section at a time, a human to verify each section against the source. You still save time, but the failure mode changes from “confident wrong answer on the whole job” to “one bad section I caught in review.”

The rule of thumb: the longer and more open-ended the job, the smaller the chunks you delegate, and the more verification you build in. GDPx does not argue against agents. It argues against unsupervised agents on expert work.

A small-business playbook for the next 90 days

If you run a small business or freelance practice, here is how to act on GDPx without overspending:

1. Pick by your domain, not the headline. Check the routing table above. If your work is mostly writing and software, the Carbon-vs-Opus gap matters less than whether you can even access Carbon — and if you cannot, Opus 5.5 or GPT-6 Astra on your specific task type is the honest comparison.
2. Budget per success, not per attempt. A $51 attempt sounds cheap until you need eight of them for one finished job. Multiply attempt cost by roughly 1 divided by the pass rate for your task type before you approve an agent workflow. If that number beats a human hour for the same job, run it. If it does not, keep the human.
3. Test on your own work before you scale. Run the same real task through two models, grade the outputs yourself, and record which one you would actually ship. Your tasks are not GDPx tasks, and your grading is the only score that pays your bills.
4. Keep humans on finance and legal outputs. GDPx’s zero scores under strict grading are the clearest warning in the whole benchmark. Use agents to organize, draft, and check finance and legal work — never to finalize it unsupervised. This is also where the “not professional advice” disclaimer earns its keep: agent output on money and legal questions needs a qualified human’s eyes, full stop.
5. Re-check when GDPx refreshes. The next version lands within three to six months. If a model jumps from 12% toward 20% on the same task set, that is a real signal to widen delegation. If scores stall, the task-size framework above stays the right approach.

What to watch next

Three things will change this picture fastest. First, Gemini 4 Carbon’s general availability — if it opens up at anything near its current efficiency, the cost-per-success table flips in buyers’ favor. Second, the next GDPx version, which will show whether the 12% ceiling is moving or stuck. Third, the seven unpublished models: if a new entrant lands near the top, the “small club” framing above needs revising. Until then, the honest summary is this: agents are genuinely useful on bounded work, genuinely unreliable on whole expert jobs, and the smartest move is to delegate in small, checkable pieces while the ceiling climbs.

For the underlying leaderboard and method, see SemiAnalysis’s GDPx launch coverage via ExplainX’s GDPx breakdown and the cost-per-success analysis at Vortx.

Frequently Asked Questions

What is the GDPx benchmark?

GDPx is SemiAnalysis’s agentic benchmark, published November 2, 2026. It scores frontier AI models on 206 long tasks across 11 high-value professional domains, with up to eight hours per task, graded mostly pass or fail. The headline score is pass@1 averaged over three attempts.

Which AI model scored highest on GDPx?

Anthropic’s Claude Opus 5.5 leads GDPx v1.0 at 12.0%, followed by Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%. No published model scored above 13%.

Why did every model score zero on finance and legal tasks?

Under GDPx’s binary pass-or-fail grading, nearly every model failed every finance and legal task outright, so SemiAnalysis reported partial-credit scores for those two domains instead. Those figures are not directly comparable with the other domains’ scores.

How much does a successful GDPx task cost?

Dividing SemiAnalysis’s published cost per attempt by the pass rate gives roughly $428 per success for Opus 5.5 ($51.41 per attempt at 12.0%), about $149 for Gemini 4 Carbon ($16.29 at 10.9%), and about $562 for GPT-6 Astra ($53.99 at 9.6%). Carbon is not generally available yet, so most buyers cannot pay its price today.

Does a 12% GDPx score mean AI agents are useless for business?

No. GDPx measures the hardest tail of expert work — multi-hour jobs with strict grading. Short, well-defined tasks score far higher in practice. The practical lesson is to delegate in small, checkable pieces and keep humans verifying expert outputs, not to stop using agents.

When will GDPx be updated?

SemiAnalysis plans new GDPx versions every three to six months. The next version will show whether the 12% ceiling is moving, and whether newly published models change the leaderboard.