Gemini 4 Neon
最大上下文
1M
最大输出
65.5K
发布日期
暂无
gemini-4-carbonGemini 4 Carbon 是 Google 推出的前沿 Gemini 4 模型,适用于编程、智能体任务和复杂推理。它支持文本、图像、音频、视频和文件输入,可通过续写实现最高 100 万输出 token 的长解码,并以低、中、高推理强度控制思考深度。
上下文窗口
1M Token
最大输出
65.5K Token
发布日期
2026年11月10日
模态
| 输入价格 | 输出价格 | 缓存读取 | 缓存创建 5m |
|---|---|---|---|
| $2/M | $10/M | $0.1/M | $0.1875/M |
推理
工具调用
Temperature 参数
附件
知识库
Endpoint 协议
Gemini 4 Carbon 面向需要持续推理和跨阶段验证的复杂项目。
能分阶段推演不同方案,在证据变化时复查假设,并通过多轮验证完善结果。
能在大型迁移、架构变更和复杂调试时,分析多个组件及其依赖关系。
综合图表、表格、示意图和正文,生成可对照源材料复核的回答。
约 100 万 Token 的上下文窗口由输入材料、推理和生成文本共同使用,可在一次请求中处理大量资料。
用 Gemini 4 Carbon 处理复杂工程变更、证据核对和具有详细规格的原型。
梳理多个包之间的依赖,分阶段迁移接口,并在提交审查前逐步运行针对性测试。
对照报告、表格和示意图中的发现,起草注明依据页码的备忘录,并标出仍待核对的差异。
用参考图和工程图制作浏览器 3D 原型,再对照需求检查尺寸、交互和渲染效果。
import OpenAI from "openai"
const client = new OpenAI({
apiKey: process.env.TOKENHUB_API_KEY,
baseURL: "https://us-api.tokenhub.com/v1",
})
const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)了解 Gemini 4 Carbon 的上下文上限、思考档位、支持的输入与 API 接入方式。
它是 Google Gemini 4 系列中面向高难度、多阶段工作的模型,适合复杂软件工程、分析写作和资料推理。
Google 文档给出的输入与输出共享上限为 1,048,576 Token,包含思考 Token。流式调用单段最多 65,536 Token,非流式单段最多 32,768 Token。支持长解码的 Google 端点可在剩余上下文内分段继续生成。
Google 文档列出的输入包括文本、图片、音频、视频和 PDF 等文档,输出为文本。具体附件格式与限制以所选端点为准。
Google 文档列出的 gemini-4-carbon 思考档位为 low、medium 和 high,默认为 medium。high 为复杂任务留出更多分析和验证空间,也可能增加延迟与 Token 用量。
模型在你的 TokenHub 目录中可用时,使用 TokenHub API 密钥和页面显示的模型 ID。已启用端点的基础 URL 和请求格式可在 API 接入部分查看。
截至 2026 年 11 月 30 日,Google API 的标准档优惠价为每百万输入、缓存输入和输出 Token 分别 2、0.10 和 10 美元。自 2026 年 12 月 1 日起,分别为 4、0.20 和 20 美元。输出包含思考 Token。TokenHub 费率以本页价格表为准。
需要更充分探索和验证的任务适合考虑 Carbon,例如多包迁移、架构问题排查或跨多份密集文档核对证据。
通过一个 API Key 接入 Gemini 4 Carbon 与更多 AI 模型。
Gemini 4 Carbon 视频与社区讨论
收录 Gemini 4 Carbon 的发布帖、实测演示和开发者讨论。
X (Twitter)
Google ·
Gemini 4 Carbon is now available. It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning. If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems. https://t.co/4NVqHaMOPT
https://pbs.twimg.com/amplify_video_thumb/2120183679506841600/img/gUP9nkAcN0y69fRe.jpg查看原文Google DeepMind ·
Start building with Gemini 4 Carbon: our strongest model for complex, long-horizon tasks. It’s designed to explore problems thoroughly and run extended verification loops across software engineering, cybersecurity, knowledge work, and science. https://t.co/avhcExdUlr
https://pbs.twimg.com/amplify_video_thumb/2120218949774098432/img/CUcpPK8n8Ew04Ma_.jpg查看原文Google ·
For your biggest builds, we're previewing /swarm in @Antigravity, turning one prompt into a coordinated team of agents you can track in real time. Assign Carbon to high-level roles like architectural design, coordination, and auditing while parallel agents build and test every https://t.co/c5IOJQy2gv
https://pbs.twimg.com/amplify_video_thumb/2120184469256650752/img/o296P3lcEnhq2sQL.jpg查看原文Reddit
Gemini 4 Carbon is live
Lberondroth ·
Thank you to everyone who has contributed to this. Happy Coding, Everyone ^^ (Only for Ultra...)
查看原文My preliminary results on Carbon
Nichiren ·
展开全文
I've been seeing questions on how Carbon performs compared to Argon and I have some preliminary results based on my early measurements on work done today if anyone's interested. Take this with a huge grain of salt until I have more samples, but it does look promising so far for my workload. Tech stack is Go, HTMX, Tailwind CSS, Templ. Report: Four Carbon phases have been audited so far: three at medium and one at high. All four landed on the first run, with no retries and no Opus takeovers, and the Opus audits barely touched them. bin/seat-report --since 2026-11-10 (today's rows): Gemini 4 Carbon, day one in our coding-agent pipeline: medium vs high (tiny sample) Setup: an unattended pipeline builds a Go/HTMX web app one phase at a time. Each phase has a written contract and runs headless on one model. Then Claude Opus 5.5 audits it at high effort and fixes whatever it finds. We score each phase by how many lines the audit had to change and by the audit's findings, sorted by severity. Effort is not randomized. Harder phases (schema, permissions, hot-path queries) run at high and routine ones at medium. So medium vs high below is also routine vs hard work. Phases audited on Carbon's first day, by model and effort Model · effort Phases Landed first try Lines changed by audit (median) Findings (high / med / low) Gemini 4 Carbon · high 1 1/1 2 0 / 0 / 1 Gemini 4 Carbon · medium 3 3/3 0 0 / 1 / 0 Gemini 4 Argon · high 3 3/3 617 2 / 8 / 7 Gemini 4 Argon · medium 1 1/1 44 0 / 1 / 2 Codex Sol · high 3 3/3 6 0 / 1 / 5 Claude Opus 5.5 · high 1 1/1 263 0 / 3 / 4 The four Carbon phases Effort What the phase was Size Run time Tokens Audit result high Data model: new columns and tables, transactional data-layer methods ~950 code + ~2,000 test lines 67 min 1.15M Passed; 1 low finding, 2 lines changed medium Port of a language-detection library from a sibling codebase 230 files 50 min 0.71M 1 medium finding (a 6-line test fix) medium Copy, i18n catalog and template leftovers ~580 code lines 66 min 1.31M Passed, nothing changed medium One shared helper replacing ~25 copies 69 files 43 min 0.53M Passed, nothing changed Caveats n = 4 for Carbon and 1 at high. This is an anecdote, not a benchmark. Effort follows difficulty, so these numbers can't tell you whether medium beats high on the same task. Different phases are different work. The Argon and Opus rows come from other plans that ran the same day. One codebase, one auditor model, and contracts written in advance, which favors models that follow a spec closely. Takeaway so far: Carbon is very clean at both efforts. Our one high-effort phase was the cleanest high-difficulty result of the day. Nothing here suggests medium is better than high. Public benchmarks (Artificial Analysis: Terminal-Bench 4.0 is 60.6% at medium vs 65.2% at high) also put high ahead on quality, with medium cheaper and faster. I'll update when there are more high-effort rows. Clarification edit: Gemini runs are through the Antigravity CLI (agy) harness.
Gemini 4 Carbon Just Redeemed Google on 3D for Me: Same Prompt, New Result
SuspiciousHunter1359 ·
查看原文YouTube
Bug Triage with Gemini 4 Hydrogen and Gemini 4 Carbon
Google Antigravity
查看原文Gemini最上位モデル『Gemini 4 Carbon』きた!徹底検証(ClaudeとGPTモデルとも比較)
KEITO【AI&WEB ch】
查看原文【待ってた】Gemini 4 Carbonを実際に使ったので徹底解説!Argonから順当進化で唯一無二の魅力も健在
まさおAIじっくり解説ch
查看原文