The best AI agent in the world scores just 12% on SemiAnalysis's new GDPx benchmark. Released on 2 November 2026, GDPx tests agentic capability across 11 professional domains, and even the leading model fails roughly 88% of its frontier tasks. That number should reset expectations for anyone buying or building an agent: full autonomy is not here yet, so buy a supervised, narrowly scoped agent with human checks.

This guide explains what GDPx measures, what the leaderboard really says, and how to use it to choose the right model and scope.

The Best AI Agent Scores Just 12%: What GDPx Means for Your Business

What GDPx is — and why it is different

GDPx is SemiAnalysis's benchmark for general agentic capabilities, published on 2 November 2026 as the successor to OpenAI's GDPval (25 September 2025). Version 1.0 contains 206 tasks, built with 14 specialised data partners and ranked across 13 models. Tasks can run up to eight hours — closer to a real professional job than a short test.

The 11 domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Government, real estate and insurance are deliberately excluded. For the basics, see our guide to AI chatbots versus AI agents.

How the headline score is calculated

The headline score is pass@1 averaged over three attempts, with every domain weighted equally. Equal weighting means a strong coding score cannot hide a weak area — which is why the numbers look low, and why they are more honest than most benchmark claims.

Why the harness matters

Models ran in native harnesses where available: Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon, with an OpenCode fallback. An agent is never just a model; the tools around it shape the result. When you commission AI agent development, you buy that whole system, not just a model name.

The GDPx leaderboard in plain English

Model	Overall score	Cost per task
Claude Opus 5.5	12.0%	$51.41
Gemini 4 Carbon	10.9%	$16.29
GPT-6 Astra	9.6%	$53.99
Meta Muse Wonder Full	9.0%	Not disclosed

Opus 5.5 leads at 12.0% and is the most token-hungry of the top three, at 548,000 output tokens per task. Gemini 4 Carbon is close behind at 10.9% but costs less than a third as much, at $16.29 against $51.41. GPT-6 Astra is third at 9.6% and $53.99; Meta Muse Wonder Full fourth at 9.0%. See Unite.AI on the GDPx launch for wider context.

The honest reading is not "Claude wins": first to fourth spans only three percentage points, while the cost gap is more than threefold. For most businesses, cost matters far more than rank.

Why finance and legal scored zero for every model

Every model scored 0% in finance and legal. GDPx tasks there are long and unforgiving: one wrong clause or misread number fails the whole task, with no partial credit.

SemiAnalysis also notes Opus 5.5 and GPT-6 Astra scores are depressed by safety classifiers in cyber, biology and alignment. Some failures are refusals, not capability gaps — and a model that refuses a risky step may look worse on a benchmark yet be safer in your business. Read a zero as "do not hand this work over unsupervised", not "never use an agent here".

Cost per task matters as much as score

At $51.41 per task, Opus 5.5 costs roughly three times Gemini 4 Carbon at $16.29, for a 1.1 point lead. Run hundreds of tasks a month and that difference decides whether a pilot pays for itself.

This is the same discipline behind custom software development and SaaS development: cost per completed job, not headline capability, decides whether a system earns its keep. See also our guide to AI agents for small business automation in 2026.

A simple cost check before you buy

• Count tasks you will actually run monthly, not ones you could run.
• Multiply by cost per task, then add review time.
• Compare that total with doing the same work manually.
• If it only wins unsupervised, it does not win yet.

Pick the model by domain, not by headline rank

The domain winners tell a more useful story than the overall table. Opus 5.5 leads in quant trading, hardware design and chip design; Gemini 4 Carbon in biology and software engineering; GPT-6 Astra in recursive self-improvement and materials science; Muse Wonder Full best on cyber. If your work sits in one domain, the overall winner may not be your winner.

That is why a single-model strategy is usually wrong: a coding-heavy team may get more from Gemini 4 Carbon, a security workflow from Muse Wonder Full. See our analysis of Gemini 4 Carbon for cybersecurity. Shortlist by domain, pilot on your own tasks, and measure pass rate and cost on your work. Most serious systems route jobs to different models under shared guardrails; if you need that, talk to a team that builds bespoke AI agents with routing and human approval built in.

What "fails 88%" should change about your buying decision

A 12% pass rate on frontier tasks does not mean agents are useless. Everyday work — answering routine questions, triaging messages, booking, chasing, drafting, summarising — is far simpler than an eight-hour chip-design task, and agents handle much of it well. The mistake is jumping from "agents are useful" to "agents can run the business alone".

Buy in this order. Pick one narrow workflow with a clear success test: did the booking land, did the invoice match? Scope the agent to that workflow only. Put a human check on every output touching money, customers or compliance until the error rate is proven low on your own data. Expand only when checks show it is safe. This staged approach sits behind our AI calling agent and restaurant ERP work.

• Do buy: a narrowly scoped agent with a clear success test and human approval.
• Do buy: a pilot measured on your tasks, your volume and your cost per completed job.
• Do not buy: full autonomy over money, legal output or customer commitments on hype alone.
• Do not buy: a headline score as a substitute for testing on your own work.

A practical GDPx-informed rollout plan

GDPx updates every three to six months, so treat it as a moving picture, not a final verdict. Keep workflows, prompts and checks portable rather than locking business logic into one model.

Step 1: Choose one workflow and one success test

Pick a measurable workflow — order support, appointment handling, lead qualification — and define success first.

Step 2: Pilot with supervision and real volume

Run the pilot on real work, not demos, with human review of outputs that matter. Track pass rate, cost per completed job and review time. The same discipline applies to ecommerce development and CRM development.

Step 3: Expand only on evidence

Widen scope when pilot data supports it: error rate low, cost per job below the manual alternative. Keep the human check longest where mistakes cost most. Our teams in the United Kingdom, UAE, USA and Pakistan apply the same staged rollout. To scope a pilot, contact Coheser.

Revisit your model choice at each GDPx update. The benchmark's real gift is a habit: measure on hard, realistic work, count the cost, and let evidence decide how much autonomy you hand over. See AI Weekly on GDPx and the 12% frontier score and our guide to Gemini 4 Argon pricing before 30 November.

Frequently asked questions

What is the GDPx AI agent benchmark?

GDPx is SemiAnalysis's benchmark for general agentic capabilities, released on 2 November 2026: 206 tasks across 11 domains, running up to eight hours, succeeding OpenAI's GDPval from September 2025.

Which AI model scored highest on GDPx?

Claude Opus 5.5: 12.0% at $51.41 per task. Gemini 4 Carbon: 10.9% at $16.29. GPT-6 Astra: 9.6%. Meta Muse Wonder Full: 9.0%.

Why did every model score 0% in finance and legal?

These tasks are long and unforgiving: one wrong figure or clause fails the whole task. A zero means supervision and human sign-off, not full autonomy.

Should my business buy an AI agent now, or wait for higher scores?

Buy narrowly now, and wait on full autonomy. A supervised agent scoped to one workflow, with human checks on money, customers and compliance, can pay for itself today.