Image: Wikimedia Commons / NASA (public domain, ILLUSTRATIVE — a humanoid robot. This is not a photo of any AI model: GDPx is a benchmark, not a device, so this post uses a freely licensed illustrative photo instead.)
TL;DR — Claude Opus 5.5 Gemini 4 Carbon GDPx benchmark results in one line: on SemiAnalysis's new GDPx agent test, Anthropic's Claude Opus 5.5 tops the table at just 12.0%, Google's Gemini 4 Carbon is second at 10.9% — but Carbon completes each task for $16.29 against Opus's $51.41, so the "best" model and the best-value model are not the same thing.
📌 KEY HIGHLIGHTS
- The winner: Claude Opus 5.5 leads GDPx v1.0 with 12.0% — the best score any frontier model managed.
- The value pick: Gemini 4 Carbon is second at 10.9% but costs roughly a third of Opus per task ($16.29 vs $51.41).
- The test itself: 206 long, hard tasks across 11 professional domains, with up to eight hours per task.
- The honest caveat: 12% is low by design — the tasks were chosen to be brutally hard, not to represent everyday office work.
- The zero scores: every model scored effectively 0% in finance and legal under strict pass/fail grading.
What is the SemiAnalysis GDPx benchmark?
GDPx is a new benchmark from SemiAnalysis, the chip and AI research firm, released on 2 November 2026 and authored by Max Kan, Andrew Megalaa and Dylan Patel. Unlike chatbot tests that grade single answers, GDPx measures agentic capability: can a model carry out long, multi-step professional work on its own, using tools, over hours — the kind of work AI companies say agents will soon do for real businesses?
The first version contains 206 tasks across 11 domains chosen for expected economic return: chip design, hardware design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, materials science, alignment and software engineering. Tasks were built with 14 specialised data partners. Each task allows up to eight hours, models run in their own native harness (Claude Code for Opus, Codex for OpenAI's models, Antigravity for Gemini, OpenCode as fallback), and the headline score averages pass-at-first-attempt results across three attempts, weighting every domain equally. The full leaderboard is public at gdpx.semianalysis.com.
Claude Opus 5.5 Gemini 4 Carbon GDPx benchmark results: the leaderboard
Here is the top of the debut leaderboard, with the average cost per task SemiAnalysis reported:
| Model | GDPx score | Cost per task | Domain wins |
|---|---|---|---|
| Anthropic Claude Opus 5.5 | 12.0% | $51.41 | Quant trading, hardware design, chip design |
| Google Gemini 4 Carbon | 10.9% | $16.29 | Biology, software engineering (narrow) |
| OpenAI GPT-6 Astra | 9.6% | $53.99 | Recursive self-improvement, materials science |
| Meta Muse Wonder Full | 9.0% | Not reported in top summaries | Best available cyber model, per SemiAnalysis |
| Anthropic Claude Fable 5.1 | 8.5% | $67.16 | — |
| Google Gemini 4 Argon | 7.3% | $12.79 | — |
Illustrative image: industrial robot arms at work. Image: Wikimedia Commons — GDPx tests AI agents on professional tasks, including hardware and manufacturing-adjacent domains.
Why the best AI agent scores just 12%
A 12% top score sounds like failure — until you see what the test is. GDPx tasks are deliberately frontier-level: long, expert, multi-hour problems in fields like chip design and quant trading, graded mostly pass-or-fail. A model either completes the whole task or it does not. SemiAnalysis frames GDPx as a true frontier benchmark precisely because even the best model clears only about one task in eight. The score is not an estimate of how often an agent fails at ordinary work; it measures how far agents still are from autonomous expert work.
The finance and legal results make the point starkly: under binary grading, basically every model scored exactly zero in both domains, which is why SemiAnalysis reports partial-credit scores there instead. And there are caveats that cut both ways: SemiAnalysis says Opus 5.5's and GPT-6 Astra's overall scores are held back by their own safety classifiers in cybersecurity, biology and alignment — on a second leaderboard stripping those three domains out, Astra edges ahead of Opus by a tenth of a point and Carbon drops to fifth. GDPx is also SemiAnalysis's own paid product, its tasks have not been independently audited, and the score measures the model's harness as well as the model. Treat the ranking as a serious but vendor-run first reading, not gospel.
Why it matters for India readers
There is no India-specific announcement here — GDPx is a global benchmark — but the cost split is the part Indian readers should watch. India's IT services firms, startups and global capability centres are among the world's biggest buyers of AI agent capacity, and they buy on cost per completed task, not leaderboard glory. Carbon's profile — 10.9% at $16.29 per task, against Opus's 12.0% at $51.41 — is exactly the trade-off an Indian engineering team faces: a slightly lower success rate at a third of the price. For biology and software-engineering workloads, where Carbon actually leads, the value case is stronger still.
The other India angle is talent. GDPx's domains — chip design, cybersecurity, software engineering — map closely onto fields where Indian engineers already work, and the benchmark's message is that agents are assistants for expert work, not replacements: even the best model needs human oversight on 88% of frontier tasks. For context, Carbon itself is still limited-availability — Google's Gemini 4 Carbon launch in late October targeted a subset of cyber defenders, with wider access promised through its Fairwind Program. The same companies are meanwhile pushing AI to everyday Indian users, as Google's Dreambeans India launch this week shows.
Should you care?
⏳ Verdict: Follow the benchmark, not the headline. Opus 5.5 is the strongest all-round agent on this test, but Carbon's cost efficiency is the more useful finding for anyone who will actually pay for agent runs. The real story is the gap itself: at 12%, frontier agents are impressive tools and unreliable experts. Buy capability per rupee, keep humans in the loop, and watch GDPx's next update in three to six months — that trend line matters more than today's ranking.
FAQs
What is the SemiAnalysis GDPx benchmark?
GDPx is a benchmark released on 2 November 2026 that tests how well frontier AI models perform long, multi-step professional tasks — agentic work, not single-question answering. Version 1.0 has 206 tasks across 11 domains, built with 14 specialised data partners, with up to eight hours allowed per task.
Which AI model won the GDPx benchmark?
Anthropic's Claude Opus 5.5 topped the debut leaderboard at 12.0%, ahead of Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%. Opus won quant trading, hardware design and chip design outright, while Carbon led biology and narrowly won software engineering.
Why is the best GDPx score only 12%?
Because the tasks are deliberately frontier-level and graded mostly pass-or-fail: a model must complete an entire expert task to score. The 12% figure measures distance from autonomous expert work, not failure at everyday tasks — and in finance and legal, every model scored effectively zero under strict grading.
Which GDPx model is cheapest per task?
Among the leaders, Google's Gemini 4 Carbon is by far the cheapest at $16.29 per task, against $51.41 for Claude Opus 5.5 and $53.99 for GPT-6 Astra. Gemini 4 Argon is cheaper still at $12.79 per task, but scores lower at 7.3%. Cost per task is a benchmark average, not a production price quote.
Takeaway: The world's best AI agent completes just 12% of GDPx's hardest professional tasks — and the second-placed model does nearly as well for a third of the price. Capability is rising; value, not just victory, is the story to watch.
Follow World Tech on TechByCountry for the next update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- SemiAnalysis — GDPx announcement and leaderboard (2 November 2026), via the firm's X post: x.com/SemiAnalysis_ and leaderboard at gdpx.semianalysis.com
- Gadgets360 — "Gemini 4 Carbon Frontier AI Model Falls Slightly Behind Claude Opus 5.5 in Agentic Capabilities, but Not in Cost Efficiency" (3 November 2026): gadgets360.com
- Unite.AI — "SemiAnalysis Launches GDPx Benchmark Spanning 11 Agentic Domains" (2 November 2026): unite.ai
- Dealroom — "SemiAnalysis launches GDPx: the best AI agent completes just 12% of its high-value tasks" (2 November 2026): dealroom.co
Comments
Post a Comment