跳到正文
The Elec· Hyun-Seon, Park·· 20 天前AI 評分74

Google Gemini 4 Carbon 登上 Artificial Analysis Intelligence Index 榜首

Google’s Gemini 4 Carbon Tops Artificial Analysis Rankings

AI 導讀

Google 的 Gemini 4 Carbon 在 Artificial Analysis Intelligence Index 取得 59 分,超越 Claude Opus 5.5 排名第一。

正文

Gemini 4 Carbon scored 59 points in AA's overall evaluation, outperforming Anthropic's Claude Opus 5.5. (Source: AA)

Gemini 4 Carbon scored 59 points in AA's overall evaluation, outperforming Anthropic's Claude Opus 5.5. (Source: AA)

Gemini 4 Carbon scored 59 points in AA's overall evaluation, outperforming Anthropic's Claude Opus 5.5. (Source: AA)
Gemini 4 Carbon scored 59 points in AA's overall evaluation, outperforming Anthropic's Claude Opus 5.5. (Source: AA)

Google, which had faced concerns that prolonged delays in releasing a frontier AI model had left it behind in the AI race, is seeking to regain ground with the successive launches of Gemini 4 Argon (Argon) and, two weeks later, Gemini 4 Carbon (Carbon).

Carbon, unveiled on Oct. 28 local time, the same day Alphabet (Google) announced its third-quarter earnings, ranked first in evaluations by Artificial Analysis (AA), an independent AI model evaluation platform.

Google describes Carbon as its most powerful model to date, with leading capabilities in cybersecurity and long-horizon task execution. Alphabet CEO Sundar Pichai also highlighted Carbon during the earnings conference call, expressing confidence that Google remains at the forefront of AI.

Carbon scored 59 points on the Artificial Analysis Intelligence Index, overtaking Claude Opus 5.5 to claim first place. It achieved the highest scores in four of the nine benchmarks used to calculate the index: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.

Google's AI coding tool Antigravity CLI showed substantial performance improvements when powered by Carbon rather than Argon. (Source: AA)
Google's AI coding tool Antigravity CLI showed substantial performance improvements when powered by Carbon rather than Argon. (Source: AA)

The AA Intelligence Index was recently upgraded to version 5. It calculates scores using nine tests across four categories: knowledge work (AA-Briefcase 15%, AutomationBench-AA 5%, GDP.pdf 5%), coding (FrontierCode 2.0 15%, Terminal-Bench 4.0 10%), science (Terminal-Bench-Science 15%, CritPt v2 hard 10%) and general capabilities (AA-Omniscience and HLE-Diamond).

Carbon also scored highly on the AA Coding Agent Index, although it fell short of Claude Opus 5.5. Nevertheless, AA described the result as a significant achievement given Google's previous relative weakness in coding performance.

Google emphasized that its internal benchmarks showed substantial improvements in secure code generation, along with lower rates of misalignment and hallucinations in real-world coding environments used by internal teams. The company also claimed that efforts to reduce reward hacking during training brought the cheating success rate on Terminal-Bench 4.0 down to 0.0% and the attempt rate to 0.4%, while accuracy improved. According to Google, high-severity misalignment alerts declined by approximately 80%.

While Carbon competes closely with Claude Opus 5.5 in performance, it is relatively less expensive.

Running the nine benchmark tests used for AA's Intelligence Index cost $7.76 per task. That was approximately 23% less than Claude Opus 5.5's $10.04 and about 1.2 times the $6.25 cost of OpenAI's GPT-6 Astra.

Google is offering a 50% promotional discount to coincide with the official launch, but AA applied standard pricing in its performance analysis.

Gemini 4 Carbon's cost per task is lower than that of Claude Opus 5.5, which offers comparable performance. (Source: AA)
Gemini 4 Carbon's cost per task is lower than that of Claude Opus 5.5, which offers comparable performance. (Source: AA)

According to AA, Carbon also used an average of 123,000 output tokens per task, 29% fewer than Claude Opus 5.5's 173,000 tokens at maximum reasoning effort.

Carbon also offers a stronger performance-to-cost ratio than Argon. The two models have identical standard pricing, and Argon's cost per task was $7.68, similar to Carbon's. However, Carbon's overall index score was nine points higher.

Like the previously released Argon, Carbon is currently being rolled out gradually to a limited group of selected users, restricting public access. Pichai said during the earnings conference call that Google is developing a distilled version of Carbon.

來源:The Elec · thelec.net