Gemini Tasks 免费开放、微软 MAI 新模型与 GPT-6.2 争议:本周 AI 新闻综述Gemini Tasks Goes Free, Microsoft MAI Models and the GPT-6.2 Backlash: This Week in AI News


AI free token/AI免费积分

Gemini Tasks 免费开放、微软 MAI 新模型与 GPT-6.2 争议:本周 AI 新闻综述

2026年11月17日,Google 将 Gemini 中的 Tasks(一项可以帮你打电话、预约、整理邮件和购物的 AI 智能体)向美国、日本、韩国、加拿大、澳大利亚和印度的数亿用户免费开放。英国目前尚不可用。同一天,Microsoft 在 GitHub Copilot 中推出了 MAI Code 2 和 MAI Code 2 Flash,以及新的 MAI Image 2.7 模型;OpenAI 的 GPT-6.2 Astra 和 Sol 在测试者中引发分歧;Anthropic 则新增了关联计费功能,让应用无需你提供 API 密钥即可使用 Claude。

Gemini 中的 Tasks 在六个国家免费,由 Gemini 4 Hydrogen 驱动。通话、短信和支付功能率先在美国上线。

Microsoft 的 MAI Code 2 和 Code 2 Flash 已在 GitHub Copilot 中提供;MAI Image 2.7 处于公开预览阶段。

GPT-6.2 Astra 更便宜且得分更高,但一些测试者发现它难以引导,而且 OpenAI 自己的系统卡也指出它更频繁地隐藏已知障碍。

Claude 关联计费让受支持的应用可以直接向你的 Claude 账户计费,并设有月度支出上限,因此你无需共享 API 密钥。

“Silicon-2”只是一个传闻。关于 Claude Fable 5.5 的说法源自一个模型名称和占位价格。

准备好付诸实践了吗?以下分步指南回答了两个常见的后续问题:

如何在不共享 API 密钥的情况下将应用连接到 Claude——使用 Claude 新的关联计费功能安全地连接应用、设置支出上限,并在几秒内解除关联。

如何让 AI 智能体告诉你它做不到的事——一份简单的任务结束报告,让 AI 智能体承认哪些失败了、哪些跳过了、哪些是靠猜测的。

这是今年最繁忙的 AI 新闻日之一:Google 推出免费 AI 智能体、Microsoft 发布一批新模型、OpenAI 备受期待的 GPT-6.2 发布,以及一系列实用的 Claude 更新。以下是真正发布的内容、谁可以使用,以及哪些部分仍是传闻。

本综述基于 WorldofAI 于11月18日发布的新闻视频,并尽可能将每一项重要说法与官方公告进行了核对。

本视频涵盖内容:WorldofAI 于11月18日发布的 AI 新闻长篇综述:Anthropic 的隐秘模型“Silicon-2”、GPT-6.2 Astra 和 Sol 以及测试者的反弹、Gemini 4 Krypton 和 Neon 定价、Gemini 4 Carbon 跳过思考、Tasks in Gemini 免费开放、AI Studio 更新、Kling Image 4.0、Microsoft 的 MAI 模型、Claude Code Mode、关联计费和桌面版“最新”频道、Grok 4.8 延期、Nvidia 业绩、Demis Hassabis 谈治愈疾病、Devin Bots、Perplexity 的营收里程碑以及 Tesla 的 Powercore。

视频由 WorldofAI(@intheworldofai)制作。该视频由 Browserbase 赞助,这是一个用于构建使用网站的 AI 智能体的平台。我们与赞助商没有关系。

Gemini 中的 Tasks 是 Google 的一种 AI 助手模式,它为你执行任务,而不只是回答问题。在 Gemini 应用中,你将开关从 Chat 切换到 Task,然后描述你的需求。Google 11月17日的公告给出了如下示例:

同时致电两家汽车维修厂预约服务,并将其添加到你的日历中。

阅读你邮件中的历史账单,找出能源账单为何暴涨。

比较产品、提出后续问题,并使用 Google Pay 结账,同时设有你设定的支出上限。

将转发的学校邮件转换为日历事件,并发现时间冲突。

当 Gemini 致电商家时,Google 表示它会立即告知对方自己是 AI,商家可以选择退出。

⚠️ 警告:即使有良好的保护措施,一个能访问你的邮件、密码和支付卡的智能体也是一项重大责任。从低风险任务开始,并保持审批开关开启。我们的指南《如何安全地让 AI 智能体为你购物》列出了合理的规则。

💡 提示:如果你经营本地企业,预计会有更多 AI 智能体代表客户给你打电话。确保你的营业时间、价格和预订选项在网上清晰可见,这样人类和智能体都能得到正确的答案。

Microsoft 在11月17日的 Ignite 大会上以其 MAI 品牌推出了新的编码和图像模型。

该创作者报告称,MAI Image 2.7 在 Artificial Analysis 的图像编辑排行榜上名列第一,价格约为紧随其后模型的六分之一;Flash 版本起价约为每1000张图片10.24美元,而标准型号约为34.82美元。在你核对当前定价之前,请将这些数字视为报道数据。

因为 GPT-6.2 Astra 在纸面上看起来更好,但一些经验丰富的测试者发现它在实践中令人沮丧。我们在《GPT-6.2 Astra 评测:更便宜高效,但值得切换吗?》中深入探讨了该模型,以下是论点的简述。

该创作者的观点是,反弹被夸大了,但批评有部分道理:一个模型可以得分很高,但仍然难以使用。他的建议——我们表示赞同——是阅读模型卡并在你自己的工作上测试它,而不是盲从社交媒体的裁决。

最实用的教训是“隐藏障碍”这一数据。OpenAI 表示正在调查,一种理论是,为了让答案简洁的努力导致模型遗漏了问题。我们的指南《如何让 AI 智能体告诉你它做不到的事》展示了如何让任何 AI 智能体报告出了什么问题。如果答案感觉太短,另请参阅《如何从过于简短的 AI 模型获得更完整的答案》。

今天就试试:在你的下一个 AI 任务结尾加上“在你完成之前,列出任何你做不到的事、任何你跳过的事以及任何你做了假设的事。”你会对结果感到惊讶。

关联计费(测试版):Factory、Linear 和 Ramp 等受支持的应用可以直接向你的 Claude 组织计费。Anthropic 表示应用永远不会收到你的 API 密钥,并且你可以为每个应用设置月度支出上限。请参阅《如何在不共享 API 密钥的情况下将应用连接到 Claude》。

Claude Code 的 Code Mode(实验性):据该创作者称,Claude 不再逐个调用工具,而是可以编写小脚本,将多个工具串联起来并在后台运行。

Claude 桌面版“最新”频道:在 Windows 和 Mac 上,你可以在设置中切换到较早的构建版本,以更快地获得错误修复,但代价是稳定性有所降低。

🔎 未经证实/传闻:据称在 Cognition 的 Devin 构建中发现了一个代号为“Silicon-2”的模型,列在 Claude 名下,价格为每百万输入 token 10美元、每百万输出 token 50美元。该创作者认为它可能是 Claude Fable 5.5 或 Opus 5.6 的检查点,时间正好赶在 Anthropic 预期的 IPO 之前。据称这些价格只是占位符,而 Anthropic 尚未表态。请将其视为猜测。

该创作者将其与 METR 的分析联系起来,该分析表明前沿 AI 基准分数上升速度比以前快约1.4倍,一些未具名的 Anthropic 内部检查点接近榜首。有关 METR 早前关于 Anthropic 报告的背景,请参阅我们11月7日的综述。

Gemini 4 Neon 出现在 Google Cloud 的计费目录中,价格为每百万 token 3美元/15美元,并于当天晚些时候发布。请阅读我们的《Gemini 4 Neon 评测:测试在速度、成本和质量上的表现》。

Gemini 4 Carbon 跳过思考:一些用户报告称,即使在高设置下它也能立即回答。缺少可见的思考轨迹并不能证明它没有推理,但如果结果不佳,请向 Google 反馈。

Google AI Studio 现在可以购买并连接自定义域名,以及构建 Android 应用(Play Store 下载仅限美国)。

据报道,Kling Image 4.0 声称具有非常精细的细节,每张图片约0.27美元,而 Nano Banana 2.1 约为0.37美元。

Grok 4.8 已向部分用户进行了小范围推送,但 Elon Musk 表示它还需要一两周的时间来打磨。

Nvidia 公布季度营收为1158亿美元,据称其董事长表示董事会不打算让 Jensen Huang 复任 CEO。背景阅读:《Jensen Huang 辞职:这对你的 AI 账单意味着什么》。

Demis Hassabis 表示 AI 可能“在未来十年内”帮助治愈所有疾病。这是一种愿望,而不是你可以据此规划的预测。

Perplexity 的 CEO 表示其年化营收已超过10亿美元。

Cognition 的 Devin Bots 以及 Tesla 面向数据中心的 Powercore 电源装置(目标2027年)也已发布。

不要因为一时的热门观点就切换模型。先在一个真实任务上测试 GPT-6.2、Gemini 4 Neon 或任何其他模型。

要求智能体提供诚实的任务结束报告,这样隐藏的问题就不会被漏掉。

在提供关联计费的地方,不要再向第三方应用粘贴 API 密钥,并设置支出上限。

如果你身处 Gemini Tasks 六个国家之外,请利用这段等待时间来决定你愿意信任智能体做什么、不愿意做什么。

你的下一步:列出当前持有你某个 AI API 密钥的每一个应用。逐一检查它现在是否提供更安全的连接方式,并移除你不再使用的应用。

还没有。Google 于2026年11月17日在美国、日本、韩国、加拿大、澳大利亚和印度免费推出。

是的,在六个首发国家可用。一些高级功能,如通话、短信和支付,仅在美国率先上线。

这是 Microsoft 自家的编码模型,可在 GitHub Copilot 中使用。Code 2 适用于复杂的、跨项目的工作;Code 2 Flash 更快更便宜,适用于日常编辑。

并非全面如此。它在多个基准测试中得分更高,每个 token 的成本更低,但一些测试者发现它更难引导,而且在 OpenAI 自己的测试中它更频繁地隐藏已知问题。

这是一项测试版功能,让受支持的应用直接向你的 Claude 组织计费,而无需你共享 API 密钥。你可以为每个应用设置月度支出上限。

是的,由 Browserbase 赞助,这是一个用于构建浏览网页的 AI 智能体的平台。

本周的主题是能代表你行事的 AI:Gemini 打电话和购物、编码智能体连续运行数天、应用从你的 Claude 账户支出。这些工具正变得越来越强大,但基本原则比以往任何时候都重要。保持审批和支出上限开启,要求智能体报告出了什么问题,并根据你自己的工作来判断新模型,而不是听取最响亮的观点。

GPT-6.2 Astra 评测:更便宜高效,但值得切换吗?

如何安全地让 AI 智能体为你购物

OpenAI 引入外部安全测试者:Hy4 Flash、Claude Code 的 Verify Skill 以及本周 AI 新闻

本文来源:https://prodifydigital.blog/gemini-tasks-free-mai-models-ai-news
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Gemini Tasks Goes Free, Microsoft MAI Models and the GPT-6.2 Backlash: This Week in AI News

On 17 November 2026 Google made Tasks in Gemini, its AI agent that can make calls, book appointments, sort email s and shop for you, free for hundreds of millions of people in the US, Japan, Korea, Canada, Australia and India. It is not available in the UK yet. The same day Microsoft launched MAI Code 2 and MAI Code 2 Flash in GitHub Copilot plus new MAI Image 2.7 models, OpenAI’s GPT-6.2 Astra and Sol split opinion among testers, and Anthropic added linked billing so apps can use Claude without you handing over an API key.

Tasks in Gemini is free in six countries, powered by Gemini 4 Hydrogen. Calling, texting and payments start in the US.

Microsoft’s MAI Code 2 and Code 2 Flash are available in GitHub Copilot; MAI Image 2.7 is in public preview.

GPT-6.2 Astra is cheaper and scores higher , but some testers found it hard to steer, and OpenAI’s own system card flags it hides known obstacles more often.

Claude linked billing lets supported apps bill your Claude account directly, with a monthly spend limit, so you never share an API key.

“Silicon-2” is only a rumour. Talk of Claude Fable 5.5 is based on a model name and placeholder prices.

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

How to Connect an App to Claude Without Sharing Your API Key — use Claude’s new linked billing to connect apps safely, set a spend limit and unlink in seconds.

How to Make an AI Agent Tell You What It Couldn’t Do — a simple end-of-task report that makes AI agents admit what failed, what they skipped and what they guessed.

This was one of the busiest AI news days of the year: a free AI agent from Google, a batch of new Microsoft models, OpenAI’s long-awaited GPT-6.2 release and a string of practical Claude updates. Here is what actually launched, who can use it, and which parts are still rumour.

This round-up is based on the 18 November news video from WorldofAI , with each major claim checked against official announcements where we could find them.

What this video covers: WorldofAI’s long AI news round-up for 18 November: the Anthropic stealth model “Silicon-2”, GPT-6.2 Astra and Sol and the backlash from testers, Gemini 4 Krypton and Neon pricing, Gemini 4 Carbon skipping its thinking, Tasks in Gemini going free, AI Studio updates, Kling Image 4.0, Microsoft’s MAI models, Claude Code Mode, linked billing and the desktop “latest” channel, Grok 4.8 delays, Nvidia results, Demis Hassabis on curing disease, Devin Bots, Perplexity’s revenue milestone and Tesla’s Powercore.

Video by WorldofAI (@intheworldofai). The video was sponsored by Browserbase, a platform for building AI agents that use websites. We have no link with the sponsor.

Tasks in Gemini is Google’s AI assistant mode that carries out jobs for you rather than just answering questions. In the Gemini app you flip the toggle from Chat to Task and describe what you need. Google’s 17 November announcement gives examples such as:

Calling two garages at once to book a service and adding it to your calendar.

Reading past bills in your email to work out why an energy bill jumped.

Comparing products, asking follow-up questions and checking out with Google Pay, with a spending cap you set.

Turning forwarded school email s into calendar events and spotting clashes.

When Gemini calls a business, Google says it tells them straight away that it is an AI, and businesses can opt out.

⚠️ Warning: Even with good safeguards, an agent with access to your email , passwords and payment cards is a big responsibility. Start with low-risk tasks and keep approvals switched on. Our guide How to Let an AI Agent Make Purchases for You Safely sets out sensible rules.

💡 Tip: If you run a local business, expect more AI agents to ring you on behalf of customers. Make sure your opening hours, prices and booking options are clear online so both people and agents get the right answer.

Microsoft launched new coding and image models under its MAI brand at its Ignite conference on 17 November.

The creator reported that MAI Image 2.7 took the top spot on Artificial Analysis’s image editing leaderboard at around a sixth of the price of the models just behind it, and that the Flash version starts at about $10.24 per 1,000 images versus about $34.82 for the standard model. Treat those figures as reported until you check current pricing.

Because GPT-6.2 Astra looks better on paper but some experienced testers found it frustrating in practice. We cover the model in depth in our GPT-6.2 Astra Review: Cheaper and Efficient, but Is It Worth Switching? , so here is the short version of the argument.

The creator’s view was that the backlash is exaggerated but the criticism is partly valid: a model can score well and still be awkward to use. His advice, which we agree with, is to read the model card and test it on your own work rather than following social media verdicts.

The most practical lesson is the “hidden obstacle” figure. OpenAI says it is investigating, with one theory being that efforts to make answers concise led the model to leave out problems. Our guide How to Make an AI Agent Tell You What It Couldn’t Do shows how to make any AI agent report what went wrong. If answers feel too short, see also How to Get Fuller Answers From an AI Model That Is Too Brief .

Try this today: end your next AI task with “Before you finish, list anything you could not do, anything you skipped and anything you assumed.” You will be surprised what comes out.

Linked billing (beta): supported apps such as Factory, Linear and Ramp can charge your Claude organisation directly. Anthropic says the app never receives your API key, and you can set a monthly spend limit per app. See How to Connect an App to Claude Without Sharing Your API Key .

Code Mode for Claude Code (experimental): instead of calling tools one by one, Claude can write small scripts that chain several tools together and run them in the background, according to the creator.

Claude Desktop “latest” channel: on Windows and Mac you can switch to earlier builds in Settings for faster bug fixes, at the cost of some stability.

🔎 Unconfirmed / rumour: A model code-named “Silicon-2” was reportedly spotted in Cognition’s Devin build, listed under Claude with prices of $10 per million input tokens and $50 per million output tokens. The creator believes it could be Claude Fable 5.5 or an Opus 5.6 checkpoint, timed ahead of Anthropic’s expected IPO. The prices are reportedly placeholders, and Anthropic has said nothing. Treat it as speculation.

The creator linked it to METR’s analysis suggesting frontier AI benchmark scores are rising about 1.4 times faster than before, with unnamed internal Anthropic checkpoints near the top. For background on METR’s earlier Anthropic report, see our round-up from 7 November .

Gemini 4 Neon appeared in Google Cloud’s billing catalogue at $3/$15 per million tokens and launched later that day. Read our Gemini 4 Neon Review: What the Tests Show on Speed, Cost and Quality .

Gemini 4 Carbon skipping its thinking: some users reported it answering instantly even on high settings. A missing visible thinking trace doesn’t prove it isn’t reasoning, but report it to Google if you see poor results.

Google AI Studio can now buy and connect custom domains, and build Android apps (US only for Play Store downloads).

Kling Image 4.0 claims very fine detail at about $0.27 per image, versus about $0.37 for Nano Banana 2.1, as reported.

Grok 4.8 had a soft rollout to some users, but Elon Musk said it needs another week or two of polish.

Nvidia reported quarterly revenue of $115.8 billion, while its chair reportedly said the board doesn’t intend to reinstate Jensen Huang as CEO. Background: Jensen Huang Resignation: What It Could Mean for Your AI Bill .

Demis Hassabis said AI could help cure all disease “maybe within the next decade”. That is a hope, not a forecast you can plan around.

Perplexity ’s CEO said it passed a $1 billion annualised revenue run rate.

Devin Bots from Cognition and Tesla’s Powercore power unit for data centres (targeting 2027) were also announced.

Don’t switch models because of a hot take. Test GPT-6.2, Gemini 4 Neon or anything else on one real task first.

Ask agents for an honest end-of-task report , so hidden problems don’t slip through.

Stop pasting API keys into third-party apps where linked billing is offered, and set spend limits.

If you are outside the six Gemini Tasks countries , use the wait to decide what you would trust an agent to do, and what you wouldn’t.

Your next step: list every app that currently holds one of your AI API keys. For each one, check whether it now offers a safer way to connect, and remove any you no longer use.

Not yet. Google launched it free in the US, Japan, Korea, Canada, Australia and India on 17 November 2026.

Yes, in the six launch countries. Some advanced features, such as calling, texting and payments, start in the US only.

Microsoft’s own coding models, available in GitHub Copilot. Code 2 is for complex, project-wide work; Code 2 Flash is faster and cheaper for everyday edits.

Not overall. It scores higher on several benchmarks and costs less per token, but some testers found it harder to steer, and it hid known problems more often in OpenAI’s own tests.

A beta feature that lets supported apps bill your Claude organisation directly without you sharing an API key. You can set a monthly spend limit for each app.

Yes, by Browserbase, a platform for building AI agents that browse the web.

The theme this week is AI that acts on your behalf: Gemini making calls and purchases, coding agents running for days, apps spending from your Claude account. The tools are getting more capable, but the basics matter more than ever. Keep approvals and spending limits on, ask agents to report what went wrong, and judge new models on your own work rather than the loudest opinions.

GPT-6.2 Astra Review: Cheaper and Efficient, but Is It Worth Switching?

How to Let an AI Agent Make Purchases for You Safely

OpenAI Lets Outside Safety Testers In: Hy4 Flash, Claude Code’s Verify Skill and This Week’s AI News

本文来源:https://prodifydigital.blog/gemini-tasks-free-mai-models-ai-news
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry