AI News

Google 发布 Gemini 4 Hydrogen,面向日常智能体工作流

Unite.AI更新于 11月13日 03:41

Google Launches Gemini 4 Hydrogen for Everyday Agentic Workflows

AI 导读

Google 于 2026 年 11 月 12 日发布 Gemini 4 Hydrogen,并开始在开发者、企业和消费级产品中全球推送,付费 API 定价为每百万输入 token 1.50 美元、每百万输出 token 7.50 美元。该模型由 Gemini 4 Carbon 衍生,定位快速执行,覆盖写代码、跑测试、验证改动、研究和工具调用智能体,支持文本、图像、音频和视频输入,知识截止到 2026 年 3 月。Google Antigravity 新增的 Orchestrator 选项让 Carbon 在只读阶段规划、Hydrogen 执行,遇到阻碍时可交回 Carbon 修复计划。

推荐理由

精选AI 评分 83

原文给出 Hydrogen 与 Carbon 的分工方式及 Orchestrator 动态切换机制,可据此判断快速执行模型在智能体工作流中的位置。

主题

标签

正文

Google launched Gemini 4 Hydrogen on November 12, 2026, beginning a global rollout of a model designed for fast coding and agentic work across its developer, enterprise and consumer products. Paid API pricing is $1.50 per million input tokens and $7.50 per million output tokens.

Hydrogen is adapted from Gemini 4 Carbon and fills a distinct role in the Gemini 4 lineup. Carbon and Argon handle deeper architectural judgment and extended reasoning; Hydrogen is intended for rapid execution, including writing code, running tests, verifying changes, research and tool-using agents. It accepts text, images, audio and video, and has a March 2026 knowledge cutoff that may be less current in some domains.

Orchestrator Splits Planning From Execution

The new Orchestrator option in Google Antigravity applies that division within a single session. Carbon can explore dependencies and produce a plan in a read-only phase; Hydrogen then edits code, updates packages and runs tests. If execution encounters friction, control can return to Carbon to inspect the state and repair the plan. Unlike a static router that commits an entire task to one model based on the opening prompt, Orchestrator can change models as the work changes.

Thinking Controls and Coding Benchmarks

The Gemini API documentation lists a rate-limited free tier and cached input at $0.075 per million tokens through implicit caching, alongside the flat paid pricing. Hydrogen’s input, thinking and output share one 1,048,576-token window. Synchronous responses support up to 65,536 tokens per hop, while background execution or client-side continuation can extend generation up to the remaining capacity in that window.

Hydrogen offers four thinking levels: low, medium, high by default, and the new xhigh. Google recommends high for most complex tasks and reserving xhigh for harder multistep problems where high falls short, noting that xhigh can overthink simpler prompts. These controls make the shared token budget consequential: reasoning consumes capacity that could otherwise hold input or final output, while higher settings also increase latency.

Google’s reported results include an Artificial Analysis Intelligence Index score of 51 at high, versus Argon’s 50, rising to 54 at xhigh. Against Gemini 3.8 Flash, Hydrogen finishes tasks 2.5 times faster at high, scores 57% on Terminal-Bench 4.1 versus 18.2%, and cuts fabricated claims by more than fivefold.

The model card supplies a material qualification: its cheating-adjusted Terminal-Bench 4.1 table lists Claude Sonnet 5.5 at 63.9%, above Hydrogen’s 57.0%. Hydrogen therefore improves sharply on Google’s previous Flash model without leading every competitor in the same coding comparison.

Global Rollout and Safety Assessments

Beyond the API and AI Studio, the November 12 rollout covers Gemini for business, the Gemini app, Google Search, Antigravity and Gemini in Sheets for eligible subscriptions. Free Gemini app users will receive Hydrogen through automatic model selection when Google considers it appropriate for a prompt. Google says the app follows tone and formatting instructions more closely, draws on uploads and connected services including Gmail, Drive, Airtable and Dropbox, and selects tools more reliably; Search gains better handling of multipart or typo-filled queries, more visual responses and expanded query fan-out for a wider range of links.

Google reports reductions of up to 81% in agentic overreach and 79% in under-service versus Gemini 3.8 Flash, together with greater resistance to indirect prompt injection. Those measures address failure modes that matter when a model can select tools and act across connected services.

The model card’s automated comparisons also record a 5.57-percentage-point increase in unjustified refusals versus Gemini 3.8 Flash. Its listed limitations include hallucinations, occasional slowness or timeouts, and potentially higher token use at elevated thinking levels.

The frontier-safety assessment in the card found no material general-capability increase over Argon. Targeted tests placed Hydrogen below the biological-information and cybersecurity critical capability levels, which mark severe-scale risk in Google’s framework, while reaching the alert thresholds in both domains. The card says fuller safety documentation is forthcoming.

Hydrogen’s broader significance is the division of labor it enables: frontier planning can be reserved for harder stages while a faster execution model handles work across widely used products. Whether that arrangement preserves the reported efficiency gains in everyday use, alongside the documented refusal and reliability limitations, is the central question for this rollout.

来源

Unite.AI