AI Ranking

Gemini 4 Carbon Dethrones Claude Opus 5.5 With a Bang

Jakob Steinschaden28. October 2026, 19:04

Bar charts of the Artificial Analysis Intelligence Index v5 and Coding Agent Index with Gemini 4 Carbon in first place ahead of Claude Opus 5.5 — Image: Artificial Analysis

Anthropic is gearing up for its stock market debut, and Google just grabbed pole position in the AI rankings.

Google has just unveiled Gemini 4 Carbon, which it calls its most capable model yet. The independent testing service Artificial Analysis ranks it first in its Intelligence Index v5 with 59 points. Until now, Anthropic’s Claude Opus 5.5 had led the index, with 58 points at launch. The timing is awkward for Anthropic: Its IPO prospectus is on file with the US securities regulator, and the stock is expected to start trading on the Nasdaq in the coming weeks.

Where Carbon Leads

Carbon sets the top score in four of the index’s nine individual tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. That puts the model ahead in complex reasoning, knowledge and scientific tasks. Compared with its predecessor Gemini 4 Argon, which launched for paying users two weeks ago, Carbon gains 9 points.

The jump in coding stands out, a discipline in which Google lagged behind for a long time. In the Artificial Analysis Coding Agent Index, Carbon scores 59 points with the Antigravity CLI, only 3 points behind Claude Code with Opus 5.5 (62). On Terminal-Bench 4.0 (64 percent) and FrontierCode v1.1 (51 percent), Carbon ranks second behind Opus 5.5 in both.

Carbon also looks good on reliability. In the updated AA-Omniscience test it scores 54 points, against 48 for Opus 5.5 and 46 for Argon. Accuracy rises by 10 points over Argon to 58 percent. The hallucination rate climbs from 4 to 11 percent, which Artificial Analysis says is still the second-lowest of all models, with only Argon lower. Claude Opus 5.5 and GPT-6 Astra both come in at 50 percent.

What Carbon Costs

Carbon supports a context window of 1 million tokens and accepts text, images, video and speech as input, with text as output. The regular price is $4 per million input tokens and $20 per million output tokens (about €3.56 and €17.80), the same as Argon. At launch, Google grants a 50 percent discount, bringing prices to $2 and $10 (about €1.78 and €8.90). According to Artificial Analysis, no end date has been confirmed. Cached input tokens cost 95 percent less.

Per task in the Intelligence Index, Artificial Analysis puts the cost at $7.76 at standard prices (about €6.90). That is roughly 23 percent less than Claude Opus 5.5 ($10.04, about €8.94), but 1.2 times the cost of GPT-6 Astra ($6.25, about €5.56). Carbon needs 123,000 output tokens per task, 29 percent fewer than Opus 5.5 (173,000).

Google’s Focus: Cyber Defense

Google itself pitches Carbon mainly as a cybersecurity tool. According to the company, it posts top scores for finding vulnerabilities (88.2 percent on SEC-bench Pro), exploiting them (54.8 percent on ExploitGym) and fixing them (71.0 percent on CWE-bench). The Chrome team reportedly received patches more than ten times better than from another lab’s frontier model. On WebDev Arena, Carbon gains 89 Elo points and lands in fourth place with 1769.

Google also promises progress on the model’s own safety: The share of successful cheating attempts on Terminal-Bench 4.0 drops to 0.0 percent, unsupported claims are cut in half compared with Argon, and Carbon was not successfully attacked in any case in Gray Swan’s prompt injection test. All of these figures come from Google.

Why the Throne Is Wobbly

The top spot comes with caveats. Carbon is available only to selected users for now, first to cyber defense teams and other partners through the Fairwind Program. Google is putting the model through the US government’s voluntary pre-release testing and plans to open it to developers, businesses and consumers as soon as possible afterward.

On top of that, the lead over Opus 5.5 in the overall index is a single point, and Anthropic’s model stays ahead in agentic coding. On the agentic knowledge work test AA-Briefcase, Carbon reaches 1609 Elo, up 126 points from Argon, but trails Grok 4.7 (1638), Muse Wonder (1659) and Kimi K3.1 Pro (1678). The index v5 was also only recently rebuilt with secret test data.

Anthropic Ahead of Its IPO

How much a ranking spot will affect Anthropic’s IPO remains to be seen. According to the prospectus, the company generated $16.3 billion (about €14.5 billion) in revenue in the first half of the year, and existing investors expect a valuation of $1.5 trillion to $1.75 trillion.

Anthropic is likely to respond to Google’s move. It is conceivable that the company will unveil another top model such as Claude 6 shortly before going public, to fire up investor sentiment once more ahead of the first trading day. There is no confirmation of this so far, and Anthropic has not commented on any such timing.