Gemini 4 Carbon 跑分逊于 Claude,胜在性价比
Gemini 4 Carbon 在基准测试中以低成本拿下高分
Gemini 4 Carbon 在基准测试中仅次于 Claude,但在性价比上胜出
Ben Schoon | 2026年11月2日 上午7:30 PT
Google 的新前沿模型 Gemini 4 Carbon 带来了一些值得欢迎的改进,虽然它未能在新基准测试中登顶,但性价比要高得多。
AI 研究与分析公司 SemiAnalysis 最近发布了其新的前沿级基准测试 GDPx,其中包含 Google Gemini 4 Carbon 模型的成绩。该排行榜涵盖了 OpenAI、Anthropic 和 Meta 等各家主要前沿模型,并详细列出了每个模型的得分与成本。
该基准测试包含 206 项任务,覆盖金融、材料科学、网络安全、软件工程、法律、生物学以及其他多个领域(主题)。
结果排名第一的是 Anthropic 的 Claude Opus 5.5,但 Google 的 Gemini 4 Carbon 紧随其后。两者之间的差距相对很小,但成本是这里最需要考虑的因素。Claude 的每项任务成本约为 51 美元,而 Gemini 仅略高于 16 美元。Gemini 也击败了榜单上的下一个模型 GPT-6 Astra——后者得分更低,而成本高于前两者,约为每项任务 54 美元。
报告指出,若去掉网络安全、生物学和对齐这几个领域,结果排序会发生变化。Claude 和 GPT 模型中的安全防护机制使 GPT-6 Astra 位居榜首,GPT 6.1 Sol 和 Claude Sonnet 5.5 的排名也超过 Gemini 4 Carbon。SemiAnalysis 表示,这一结果更符合他们对各模型的“主观印象”。
至于 Google 现有的前沿模型 Gemini 4 Argon,它在两份榜单中都位列第八,超过 Meta 的 Muse Wonder,但低于 Wonder Full。在这项特定基准测试中,Argon 的成本仅略低于 Carbon。
在聚焦网络安全时,Google 的两款模型得分均优于 OpenAI 和 Anthropic 的模型,但确实落后于 Meta 的 Muse 系列模型——该基准测试认为 Muse 系列“以明显优势”表现最佳。
Google 于 10 月 28 日发布 Gemini 4 Carbon,并宣称取得“显著提升”,但公开发布日期尚未确认。对大多数客户而言,它是“即将推出”。在下方 Google 自家基准测试中,它在许多方面超过了 GPT-6 Astra 和 Opus 5.5。
Ben 是 9to5Google 的高级编辑。
可在 Twitter @NexusBen 找到他。爆料可发送至 schoon@9to5g.com,或加密发送至 benschoon@protonmail.com。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Gemini 4 Carbon lags just behind Claude in benchmark, but wins in value)
Gemini 4 Carbon scores high with low costs on benchmark
Gemini 4 Carbon lags just behind Claude in benchmark, but wins in value
Ben Schoon | Nov 2 2026 - 7:30 am PT

Google’s new frontier model, Gemini 4 Carbon, brings some welcome improvements and while it doesn’t top the charts in a new benchmark, it does provide much more bang for your buck.
SemiAnalysis, an AI research and analysis firm, recently published its new frontier-level benchmark, GDPx which includes numbers for Google’s Gemini 4 Carbon model. The leaderboard includes every major frontier model from the likes of OpenAI, Anthropic, and Meta and details both the score and cost for each model.
The benchmark includes 206 tasks across finance, materials science, cybersecurity, software engineering, legal, biology, and several other domains (topics).
Topping the results is Anthropic’s Claude Opus 5.5, but Google’s Gemini 4 Carbon is very close behind. The difference between the two is relatively minimal, but the cost is the biggest factor to consider here. Where Claude’s cost came out to around $51 per task, Gemini was just over $16 per task. Gemini also beat the next model on the list, GPT-6 Astra, which scored lower while costing more than both at about $54 per task.
Advertisement - scroll for more content
The report notes that taking out cybersecurity, biology, and alignment domains shuffles the results. Safety guardrails in Claude and GPT models puts GPT-6 Astra at the top of the list with GPT 6.1 Sol and Claude Sonnet 5.5 outranking Gemini 4 Carbon. SemiAnalysis says that this better aligns with their “subjective impressions” of each model.
As for Google’s existing frontier model, Gemini 4 Argon, it ranks in eighth place on both lists, beating Meta’s Muse Wonder, but below Wonder Full. Argon’s costs in this particular benchmark are only slightly lower than Carbon.
Both of Google’s models score better than OpenAI and Anthropic models when focused on cybersecurity, but do fall behind Meta’s Muse models, which the benchmark found to be the best “by a meaningful margin.”
Google announced Gemini 4 Carbon on October 28 and touted “significant gains,” but the public release date has not been confirmed just yet. It is “coming soon” for most customers. Google’s own benchmark, below, beat GPT-6 Astra and Opus 5.5 in many aspects.
Ben is a Senior Editor for 9to5Google.
Find him on Twitter @NexusBen. Send tips to schoon@9to5g.com or encrypted to benschoon@protonmail.com.
原文为英文,正文由自动抓取并保留原文。
原文发布于 9to5Google(Gemini 4 Carbon lags just behind Claude in benchmark, but wins in value),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。