Tech • AI • Robotics • Game
Google's New CARBON is Now the Most Powerful AI in the World
TL;DR
Google has reclaimed the top spot on major AI rankings with Gemini 4 Carbon, but the model is initially restricted to a small group of cybersecurity defenders.
KEY POINTS
Top ranking in AI benchmarks
Gemini 4 Carbon was announced on October 28 and quickly rose to No. 1 on the Artificial Analysis Intelligence Index, scoring 59 against 58 for Claude Opus 5.5 and 57 for GPT-6 Astra. That marked a sharp jump from Gemini 4 Argon, which scored 50 only two weeks earlier. The index aggregates nine tests, and Carbon led in four, including reasoning, scientific research, knowledge, and automated tasks.
Google’s strongest showing in months
The release ended a roughly half-year stretch in which OpenAI and Anthropic had largely outpaced Google in public perception and benchmark momentum. Google also reported Carbon finishing first in 12 of 14 core benchmark tables and 15 of 17 across a broader internal set. On the VAL index it reached 70.4%, and on the blind-vote Arena text leaderboard it hit 1,528, also placing first.
Major gains in long-context reasoning
Carbon can process about 1 million tokens, roughly 1,500 printed pages, across text, images, speech, and video. On a long-context retrieval test, it scored 96.2%, compared with 84.2% for Argon, 71.8% for GPT-6 Astra, and 66.8% for Claude Opus 5.5. The result suggests an unusual ability to maintain coherence across very large document sets.
Science and coding performance improved sharply
On Raman Bench, a science and math test, Carbon scored 80%, well ahead of rivals. On HLE Diamond, a broad advanced-subject exam, it posted 69%, narrowly leading Opus 5.5. Coding, historically a weaker area for Google models, also improved: on Terminal Bench 4.0, Carbon rose to 65.2% from Argon’s 57.1%, though Claude and GPT still lead on some software-engineering tests such as Frontier SWE2 and OS World.
Agentic tasks are a standout
Carbon appears especially strong when allowed to work autonomously over long tasks. On Vending Bench 2, a simulated long-run business management test, it scored 1.8 times the result of Claude Opus 5.5. On Arena’s agent leaderboard it ranked second globally, behind only Opus 5.5 Max, while leading on adaptability when users change instructions mid-task at plus 18.6%.
Speed and verbosity remain drawbacks
Testers reported Carbon often taking two to four times longer than Argon to think before answering. Artificial Analysis found output speed near 100 tokens per second, but the model consumed far more total tokens than peers, generating 200 million across tests versus a typical 64 million. Average task time was about 21 minutes, and on some legal research tasks it averaged nearly 52 minutes.
Hallucination picture is mixed
On Artificial Analysis’s knowledge test, Carbon’s accuracy rose from 48% to 58%, but its hallucination rate increased from 4% to 11%. Even so, that was still the lowest hallucination rate among leading models on that chart, with Claude Opus 5.5 and GPT-6 Astra near 50%. On a combined score rewarding accuracy while penalizing fabricated claims, Carbon still led 54 to 48 over Opus.
Cybersecurity is the reason for restricted access
Google is limiting initial access because Carbon sets new highs on offensive and defensive cyber tasks. It scored 88.2% on finding real software vulnerabilities, 54.8% on determining whether they can be exploited, and 71% on writing fixes. Google Chrome engineers said it was more than 10 times better at producing high-quality security patches than another frontier model, while Wiz said it leads website-hacking evaluations in its Scan for Good project.
Safety and behavior controls improved
On Grey Swan’s prompt-injection test, Carbon posted a 0% successful attack rate over up to 15 tries, improving on Argon’s 0.7% and outperforming Opus 5.5 at 1% and GPT-6 Astra at 8.5%. On Terminal Bench 4.0, successful cheating fell to 0%. In internal coding-team use, serious misbehavior alerts dropped by nearly 80%, while overstepping fell 26% and under-delivery fell 29%.
Pricing arrives amid wider pressure on Alphabet
Carbon launches at $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%. After the introductory period, pricing rises to $4 and $20, matching Claude Opus 5.5. The launch came just before Alphabet’s quarterly earnings call, where free cash flow showed a loss of just over $9 billion for the quarter, the second consecutive negative quarter, and the stock dropped more than 10% in after-hours trading.
CONCLUSION
Gemini 4 Carbon has given Google a measurable lead in top-end AI capability, particularly in long-context reasoning and cybersecurity. The key unanswered question is whether Google can turn that benchmark win into broad market impact before competitors release their next models.
Ask a question