#产品更新

卧槽这谷歌波是真掀桌了， 他们刚出了个 Mac 端的 AI 笔记应用，免费，没有任何订阅。 叫做AI Edge Foresight，录音，转写，整理，提问，全在 Mac 本机离线跑，一分钱不收。 开会它帮你录音转写，自动整理成笔记，会后还能直接对着你所有旧记录提问。 最狠的是全程在本机跑，断网照用，录音和笔记全留在你电脑里，不上传云端。 它本地跑了两个谷歌模型，EmbeddingGemma 2 专门翻你所有记录做语义检索， Gemma 4 负责生成回答，检索加生成一条龙，全在苹果芯片上跑完，不联网。 等于 Granola 那种按月付费、录音还得传上云的活儿，谷歌免费，离线，在你自己机器上给办了。 现在还是 labs 实验版，只认苹果芯片 Mac，Intel 用不了，也不在 App Store，下载地址放评论区。

Hi! I work on Cowork. Probably no surprise, but we want to make our users as powerful as possible with Claude - and you're right that explaining and keeping up with the capabilities is hard. Affected users got both an email and an in-app notice, but I'm grateful for the feedback - I think this change was more surprising to people than I'd like it to be. This one will be a longer explainer, bear with me! The "old" version of Cowork runs model inference in the cloud, executing tool calls in an Anthropic-provided VM we shipped to your computer. We added the VM for capability, safety, and security reasons - mapping in just the data you explicitly added to your session. People loved what they were able to do with Claude but didn't love the disk, battery, and performance cost of running the VM locally. Also, people didn't love that closing your laptop means the work stops. The "new" version of Cowork runs model inference and the VM in the cloud. Each session gets its own sandbox, not sharing state with other sessions. When the VM needs something on the users' device (like a file), the desktop app is responsible for that file access tool call. To be clear, Claude will only access files in folders that the user explicitly added to the session, the same way Claude only saw files explicitly added to the VM in the "old" version. Sandboxes are destroyed when the session ends! We think this solves a lot of problems we've heard about (like using Cowork from a phone, keeping work running, or getting all the same power without losing battery to the VM) - but we'll keep iterating on new feedback we hear!

I literally can't think of a better risk/reward bet right now than building a ChatGPT plugin and getting in front of their 1.2 billion weekly users. As of yesterday, ChatGPT recommends plugins in the middle of conversations. Millions of questions get asked every day with no plugin to recommend. Each one is an empty slot. Facebook apps in 2007. iPhone apps in 2008. TikTok in 2019. Every big platform has a few months where distribution is free, and most people miss it. I saw this firsthand when I lived in SF building venture backed companies. They are sacred times! If I were starting this week: 1. Find a tool or workflow people already Google for or do by hand every week (PDF to Excel, invoice generator, weekly client reports) 2. Build it as a ChatGPT plugin 3. Write the description using the exact phrases people type into ChatGPT 4. Ship 5 of them and see which ones get picked up 5. Double down on the one with the most traction In 12 months these slots will be crowded but right now they're pretty empty. Most people will realize ChatGPT plugins were the biggest free distribution channel of 2026 sometime in 2027.

Ultrafast is our fastest way to build with Astra yet: in Codex, it runs up to 8x faster than Astra Standard and 4x faster than Astra Fast. Bring your ideas to life as fast as you can type them.

In our internal evaluations, Offloop achieved new state of the art results on three benchmarks evaluating professional knowledge work: 84.9 on GDPval, 44 on GDP.pdf, 67.2 on JobBench - each ahead of Codex and Claude Code, at 1/3 to 1/10 of their cost per task. Lower cost per completed task means agents you can leave running, more tasks per dollar, and margins that survive scale.

chatgpt sites is a terrible name for all the features it brings. It should've been ChatGPT Apps (if that wasn't taken haha) When you tell someone about Sites they think of a landing page hosting service What it actually is: - Full cloudflare worker - Full SQL database for storing data - Full S3 storage for files - Has environment secret you can set - Has the ability to run a full backend if needed - And run your hosted landing page as a side hustle You can build your complete apps on this... AND set custom domains AND your agent can literally build it fully for you

5.6 Terra high is underrated. Switched @clawsweeper (GitHub review bot) to it and it's ~40% faster overall with negligible quality loss. Better than 5.5 on all counts. Massively cheaper. (Tried xhigh but that negates perf wins, didn't make a noticable difference in review evals)

I've literally talked to F100 companies about how confusing skill sharing is in Claude and it..... is solved/improved by Codex and buried in a tweet thread? 🤯

Where could we improve Composer 2.5? We're working on the next model and would love your feedback. Lots of work to do (our CursorBench evals below) in the coming weeks!

Happy Friday! We've reset everyone's 5-hour and weekly rate limits.

Hosting a session next Wednesday (5/13) with the OpenAI Forum on why Codex matters beyond code. Join for the livestream and Q&A if you’re interested in the history of Codex, what we're pushing on next and see some cool use-cases.

Codex goal feature seems cool Looks like you can give Codex a goal and it’ll continue to work, plan, and test until it’s done? I’m just reading the commits here but that’s what I think it is?

Fun fact - if you have a recent commit that mentions OpenClaw in a json blob, Claude Code will either refuse your request or bill you extra money. This is an empty repo, I'm just calling Claude Code directly. Insanity.

GPT-5.5 is the first model in ChatGPT to which I’ve been able to just describe my desired response style, have it understand and remember it, and then actually respond in that exact way. I want it to write like a normal person, to default to narrative with some personality, to use minimal staccato-style lists and headers and bolding, but to feel free to use formatting when it will improve readability and scannability or help it emphasize key points. But sparingly! I previously crafted custom instructions for this with painstaking experimentation and it still didn’t work great because the models always went too far one way or another. Either too much formatting or no formatting at all, the latter of which resulted in weird things like bulleted lists that just didn’t have any bullets. Yesterday I spent time just describing to ChatGPT my desired response style, and even my background (magazine journalism, long ago, which influences my taste in writing). It understood, recorded a memory, and now writes as desired. I’ll continue to monitor and probably tweak, but right now, this is working well. (PS: One reason I did this was because I had to override outdated formatting guidance it still remembered from my instructions for older models. The new memory of how I want it to write does seem to be counteracting that older memory until it fades away, thankfully.)