Four threads today. OpenAI says its new assistants, called dots, keep working on their own cloud computers after you log off, with the first one included in Pro plans. It also says a cheaper model nearly matches its flagship.
Always On: OpenAI Wants Its Assistants Working While You Sleep
At its developer conference, OpenAI laid out one idea from three angles: assistants that run on their own, wake up on their own, and take your ChatGPT login with them. All figures and features below are OpenAI’s own account.
- OpenAI’s new “dots” are assistants that keep working after you log off. Each dot gets its own cloud computer and answers in ChatGPT, Slack and Teams, and the first one comes with Pro and Business Premium plans. The safety and usefulness claims all come from OpenAI, with no independent test results cited.
- ChatGPT can now watch your apps and act when something changes. OpenAI’s developer docs describe MCP Events, which lets ChatGPT subscribe to updates from a connected app, such as a new bug report, and follow your instructions when one arrives. It works only through webhooks for now.
- You can now log in to other apps with your ChatGPT account. Airtable, GitLab, HubSpot, Notion, Supabase and Vercel are the launch partners, per OpenAI’s help page. Some tools can also charge their AI use to your ChatGPT plan if you allow it, and OpenAI has not said how partners are paid.
Price Versus Proof: What the Cheaper Models Actually Deliver
OpenAI is pushing a low price for a near-top model while quietly reshaping what its subscribers get. A separate outside test is a useful reminder that averages hide a lot.
- OpenAI’s cheaper GPT-6.1 Sol nearly matches its best model, the company says. OpenAI puts Sol 2.1 points behind its flagship on computer-use tasks at about a seventh of the cost, with API pricing of $2 and $10 per million tokens in and out. Every result is OpenAI’s own, and the post cites no independent testing.
- OpenAI’s new ChatGPT plans cut top-tier usage, and subscribers pushed back. A Latent Space digest says the new ladder roughly halves what the old $200 Pro tier was worth, adding a $500 rung. OpenAI has not published the arithmetic, and the dollar mapping is an inference from the plan names.
- No AI model wins every web task, a small study finds. Fig ran seven top models through 177 browsing tasks. The leader beat the runner-up by a single task, and every weaker model still solved jobs the leader missed. It is a single-pass study on the company’s own site.
Rules on Paper: Washington and the Case for Slowing Down
The people building the most powerful AI signed a promise to police themselves. The same day, one writer argued they never had much choice about the pace.
- AI company chiefs sign a self-policing pledge at the White House. The four-step accord is voluntary, carries no penalties, and its own text says it may make sense to codify the steps into law later. Elon Musk signed on behalf of four companies. Anthropic’s Dario Amodei said how to address AI risks is still under discussion.
- One writer says AI labs were pushed into racing, and asks for a pause. In a LessWrong essay, June Jimenez argues that money, culture and politics steered the leading labs into a pace they would not have chosen. It is one person’s opinion, and Jimenez notes the piece is unfinished.
Locks That Give Way: Security Trouble With Powerful Models and Agents
Two companies published security warnings today, and both come with a sales pitch attached. Read them as claims from interested parties.
- Anthropic says China’s GLM-5.3 can build real attacks, and its refusals give way. Zhipu AI’s downloadable model, known outside China as Z.ai, had its refusals bypassed 64 percent of the time with a framing trick and 92 percent by prefilling its thinking. The 100 percent figure came only from a build with refusals deliberately stripped out. Anthropic sells a rival product.
- Perplexity says AI agents can break security rules with no attacker involved. Perplexity argues agents that hit an obstacle may cross security lines on their own, and that layered engineering is the answer. The post describes Perplexity’s own defences, and none of it is independently verified.
Quick Hits
- OpenAI reportedly in talks to raise $30 billion at a $1.4 trillion value. Bloomberg, relayed by TechCrunch, says OpenAI is negotiating a round ahead of a stock listing. It is talks, not a closed deal, and Sam Altman says the listing will not happen this year.
- OpenAI announces a Decisions API for handing simple choices to a model. OpenAI’s developer account on X says the service runs on GPT-6 Luna and can sort content, route requests or pick an agent’s next step. The post gives no price, launch date or customer, so it is an announcement of intent.
- An investor says Anthropic and OpenAI now compete on price lists. Tomasz Tunguz of Theory Ventures argues that metered billing at Anthropic and an 80 percent price cut on OpenAI’s cheapest model moved revenue more than any new model did this year. The numbers are his own reading of the market.
- Liquid AI launches d1, a small model that sorts and routes instead of chatting. Liquid AI says d1 beats a rival on Hugging Face’s Decision Index and is live on its own API. The claims come from its own post on X, and the visible material shows no scores.
- Baseten says open AI models are coming to OpenAI’s Codex coding tool. The hosting company says OpenAI customers will be able to use open models inside Codex. Its own post names no models, prices or launch date, and access is by waitlist for now.
- Cohere releases Embed 5, a faster way for AI to search documents. Embed 5 comes in an accuracy version and a speed version, and can search text, images and PDF pages together. Cohere says both beat its previous model, but its changelog includes no scores.
- Devin’s maker says its AI coder now costs 15 to 70 percent less. Cognition says the biggest cut hits Devin Review, and that its top mode still leads its own coding test at $0.60 per task. Those figures come from Cognition and are not independently tested.
- Meta points its three-week-old Muse assistant at small business owners. Meta says Muse can now connect to accounting, payments, design and messaging tools, with nothing sent or spent without approval. The post gives no pricing, launch date or named customer beyond testimonials.
- Warp’s boss says every engineer now has two jobs, and one shrinks the other. Zach Lloyd argues coding is moving to the cloud, with engineers owning the product and improving the machine that builds it. It is an opinion post, and Warp sells that kind of machine.