Google DeepMind 发布了 Gemini 4 Neon,这是它在七周内推出的第四款模型。Neon 的速度几乎是其兄弟模型 Carbon 的两倍,在大多数基准测试中表现优于 Carbon,且成本显著更低。Google 继续以惊人的速度发布产品。
Google DeepMind 已发布 Gemini 4 Neon,这已是其在七周内的第四款 AI 模型。Neon 提供一百万 token 的上下文窗口,支持文本、图像、音频和视频,Google 将其定位在 Carbon 和 Hydrogen 之间。

Google 的 Gemini 4 Neon 在 Artificial Analysis Index 中排名第三。
在其最高推理级别“xhigh”下,Neon 在 Intelligence Index 上得分为 61 分,略高于 Gemini 4 Carbon 的 60 分,位列第三,仅次于 OpenAI 的 GPT-6.2 Astra 和 Anthropic 的 Claude Fable 5.2,而 Google 在自己的对比表中省略了这两款模型。

在其自己的概览中,Google 仅将新模型与 Claude Opus 5.5 和 GPT-6.2 Sol 进行了对比。| 图片来源:Google
成本优势来自列表价格,而非 Token 效率
通过 Google AI Studio 和 Gemini for Business,每百万输入 token 的价格为三美元(缓存后为 0.15 美元),每百万输出 token 的价格为十五美元。这比 Carbon 常规价格的四美元和二十美元低 25%,但比 Carbon 的促销价格高 50%,后者有效期至十一月底。
根据 Artificial Analysis 的数据,Neon 在 xhigh 级别下的每项任务成本为 6.44 美元,比 Carbon 未折扣的 7.76 美元低约 17%。实际上,Neon 每项任务的输出 token 使用量比 Carbon 多约 15%,因此节省完全来自于更低的每 token 列表价格。在第二高的“high”级别下,Neon 达到了与 Carbon 相同的智能得分,同时将每项任务的成本减半。基准测试的处理时间也从大约 36 分钟降至约 21 分钟,因为 Neon 每秒生成约 110 个 token,几乎是 Carbon 57 个 token 速率的两倍。
编程代理成本减半,视觉分析能力强劲
该模型在 Coding Agent Index 上尤其具有竞争力。在 Google Antigravity 内部,Neon 在 xhigh 级别下得分为 59 分,击败了 Carbon(55 分)以及 Claude Fable 5.2 和 Claude Opus 5.5(均为 58 分),且价格仅为它们的一小部分。Neon 的平均编程任务成本为 6.86 美元,而 Opus 为 12.61 美元,Fable 为 15.47 美元。据 Google 称,在 SWE-sweep 测试中,Antigravity 环境使 Neon 的命中率提高了三倍,达到 39.0%。

Neon 位于 Hydrogen 和 Carbon 之间,尽管它在几乎所有基准测试中都击败了 Carbon。| 图片来源:Google
Artificial Analysis 测量到视觉处理方面有巨大提升。在文档测试 GDP.pdf 中,Neon 在“high”级别下取得 33.4% 的最高分,领先 Carbon 五个点,尽管它在“xhigh”级别下奇怪地回落至 28.0%。在 AA-Briefcase 测试中,Neon 以 1771 Elo 分领先,成为最佳演示质量模型,超过 Opus 5.5。
在 Google 自己的评估中,Neon 还在 Vals Index、Zapier 的 AutomationBench(56.3%)以及科学测试如 RiemannBench(84.0%)的各个类别中名列前茅。对于安全专业人员,Google 正通过 Fairwind Program 推出一个网络变体,该变体在 SEC-bench Pro(90.5%)和 WebExploitBench v1.2(51.26%)上获得最高分数,且价格仅为竞争对手的一半。根据 Neon(非常简短的)Model Card,该模型在 Gray-Swan 基准测试中阻止了 100% 的间接提示注入攻击。
无炒作的人工智能新闻 – 由人工策划
订阅 THE DECODER,享受无广告阅读、每周 AI 通讯、我们每年六次的独家“AI Radar”前沿报告、完整档案访问权限以及我们的评论区访问权限。
立即订阅
查看原文
Google DeepMind has released Gemini 4 Neon, its fourth model in seven weeks. Neon is nearly twice as fast as its sibling Carbon, beats it on most benchmarks and costs significantly less. Google continues to ship at a remarkable pace.
Google DeepMind has released Gemini 4 Neon, already its fourth AI model in seven weeks. Neon offers a one-million-token context window for text, images, audio and video, and Google positions it between Carbon and Hydrogen.

Google's Gemini 4 Neon ranks third on the Artificial Analysis Index.
At its highest reasoning level, "xhigh," Neon scores 61 points on the Intelligence Index, edging past Gemini 4 Carbon at 60 points and landing in third place behind OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2, both of which Google omits from its own comparison table.

In its own overview, Google only compares the new model against Claude Opus 5.5 and GPT-6.2 Sol. | Image: Google
The Cost Advantage Comes from List Price, Not Token Efficiency
Through Google AI Studio and Gemini for Business, one million input tokens costs three dollars (or $0.15 when cached) and one million output tokens costs fifteen dollars. That is 25 percent below Carbon's regular price of four and twenty dollars, but 50 percent above Carbon's promotional price, which remains in effect through the end of November.
According to Artificial Analysis, per-task costs for Neon at xhigh come in at $6.44, roughly 17 percent below Carbon's undiscounted figure of $7.76. Neon actually uses about 15 percent more output tokens per task than Carbon, so the savings come entirely from the lower per-token list price. At the second-highest level, "high," Neon matches Carbon's intelligence score while cutting per-task costs in half. Processing time for the benchmark also drops from about 36 minutes to roughly 21 minutes because Neon generates around 110 tokens per second, nearly double Carbon's rate of 57 tokens per second.
Half the Cost for Coding Agents and Strong Visual Analysis
The model is especially competitive on the Coding Agent Index. Inside Google Antigravity, Neon at xhigh scores 59 points, beating Carbon (55 points) as well as Claude Fable 5.2 and Claude Opus 5.5 (58 points each) at a fraction of the price. An average coding task costs $6.86 with Neon, compared to $12.61 for Opus and $15.47 for Fable. In the SWE-sweep test, the Antigravity environment triples Neon's hit rate to 39.0 percent, according to Google.

Neon slots in between Hydrogen and Carbon, though it beats Carbon on nearly every benchmark. | Image: Google
Artificial Analysis measured large gains in visual processing. On the document test GDP.pdf, Neon at "high" hits the top score of 33.4 percent, five points ahead of Carbon, though it oddly falls back to 28.0 percent at "xhigh." On the AA-Briefcase test, Neon takes the lead for best presentation quality with 1771 Elo points, ahead of Opus 5.5.
In Google's own evaluations, Neon also leads individual categories on the Vals Index, Zapier's AutomationBench (56.3 percent) and science tests such as RiemannBench (84.0 percent). For security professionals, Google is launching a cyber variant through the Fairwind Program that delivers top scores on SEC-bench Pro (90.5 percent) and WebExploitBench v1.2 (51.26 percent) at half the price of competing models. According to Neon's (very brief) Model Card, the model blocks 100 percent of indirect prompt-injection attacks on the Gray-Swan benchmark.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
原文版权归 The Decoder 及原作者所有,本站转载仅供阅读。如需删除,请联系 support@aisa.one。