The Wire · OpenAI
ChatGPT, the GPT model line, and OpenAI's product and research moves — tracked for builders as they ship.
OpenAI cut GPT-5.6 Luna API prices by 80% to $0.20 per million input tokens and $1.20 per million output, while Terra fell 20% to $2 and $12. Fast mode gives Sol up to 2.5 times Standard speed at twice the price. OpenAI says Sol-assisted kernel work lowered serving cost by 20% and improved token-generation efficiency by more than 15%, while Luna delivers year-old frontier performance at roughly six cents per task-dollar and nearly nine times the speed. The edition connects cheaper models to Amazon's reported $1.8m, 860%-over-budget coding task, Gemini Robotics 2 whole-body control, Nscale's Anyscale acquisition and Okta's roughly $200m Permiso deal.
If you thought Windows Recall was creepy, buckle up. Microsoft is about to shoehorn even more AI into its desktop operating system, giving Copilot access to every file on your PC. But don't worry, Redmond insists it's totally secure and safe since the agents powering the functionality will run entirely on your PC — some of the time. Microsoft describes the new functionality, announced alongside a new Nvidia N1X powered Surface Laptop, as "hybrid intelligence." Built atop Microsoft's HydraFusion model router, Windows apps and Copilot features will soon be able to use local models when it makes sense and cloud-based models when it doesn't. For example, Windows might use a local model for simple or privacy-sensitive tasks, or decide complex or performance-sensitive requests need OpenAI's GPT-6 or Anthropic's Claude. Local agents will run in Microsoft's Execution Containers (MXCs), agent sandboxes with fine-grained permissions applied at runtime. According to Microsoft EVP of Windows and Devices Pavan Davuluri, MXCs mean Windows is now "the most secure platform for agents." (That claim will surely age well because models never escape their sandboxes and routers never misbehave.) To support this hybrid intelligence functionality, Microsoft is extending support for Windows ML, its inference framework built on the ONNX runtime, to Llama.cpp – one of the highest-performing inference engines for local AI inference. By adding support for Windows ML, Llama.cpp should be able to take advantage of a broader range of accelerators. Microsoft is rolling out its hybrid intelligence functionality to GitHub Copilot — its AI code assistant — next week and its Copilot app over the next few months. The new Copilot features look and sound suspiciously like OpenClaw in a business suit. Copilot's agentic functionality will not have access to your entire file system, opening the door to advanced search functionality and productivity tasks through integrations with Microsoft Office. In a demo
Read full story →<p>As previously <a href="https://simonwillison.net/2026/Sep/28/claude-sonnet-5-5/">promised</a>, here's Anthropic's new fast, low cost model: <a href="https://www.anthropic.com/claude-haiku-5-5">Introducing Claude Haiku 5.5</a>.</p> <p>The previous Haiku, 4.5, was very much showing its age. It came out <a href="https://simonwillison.net/2025/Oct/15/claude-haiku-45/">almost a year ago</a>, and was priced at $1/million input and $5/million output - relatively expensive even back then, and a full 10x the price of OpenAI's <a href="https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/#gpt-6-sol-and-luna-are-half-the-price-of-their-gpt-5-6-equivalents">GPT-6 Luna</a>, released last month.</p> <p>The new Haiku exactly matches the price of GPT-6 Luna - $0.10/$0.50 - up to 100,000 tokens. Beyond 100,000 tokens the price increases 5x to $0.50/$2.50. Luna itself has a price increase at 272,000 tokens but only to $0.20/$0.75.</p> <p>Haiku 5.5 also uses a new, less generous tokenizer. My <a href="https://tools.simonwillison.net/claude-token-counter">Claude Token Counter</a> tool shows that the same long prompt uses around 1.25x as many tokens with Haiku 5.5 compared to Haiku 4.5, so there's a hidden price increase there.</p> <p>If your workloads fit in 100,000 tokens, Haiku is the same price as Luna and reports higher benchmark scores. Above 100,000 tokens, Luna looks like a much better deal.</p> <p>The <a href="https://github.com/simonw/llm-anthropic/releases/tag/0.30">most recent release of llm-anthropic</a> finally fixed it so I don't need to ship a new version of that plugin for every new model. I tested the new model like this:</p> <pre><code>llm install -U llm-anthropic llm anthropic refresh llm -m claude-haiku-5.5 "Generate an SVG of a pelican riding a bicycle" -o thinking_effort low </code></pre> <h4 id="pelicans">Pelicans</h4> <p>Here are <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F485dfa24b4efa126
Read full story →When OpenAI introduced dots last month, it pitched them as "always-on agents" running in their own cloud computer, but that may be overstating the readiness of the technology and of the people trying to use it. Presently, multiple users are reporting task failures and connection issues with their dots. Others have flagged cloud connectivity problems, dots that disappear, and dot refusals that appear to be related to abuse prevention. And the OpenAI Developer Community forum shows several similar concerns. OpenAI did not respond to a request for comment about the usage of dots. ABI Research argues that "AI claws" – always-on agents – will shape enterprise AI adoption. The biz sees "a market evolving from isolated assistants toward modular agent ecosystems built for longer-running, enterprise-grade automation." Despite the hiccups that have afflicted dots, some skepticism on social media forums, speculation that OpenAI rushed dots out in response to Meta Muse and Grok Bot, and security issues, dots have some enthusiastic users in addition to detractors who just don't see the point. They also have competition from companies and developers who believe they can deliver a similar service more affordably, with fewer restrictions. A company called Tawkit that makes an SDK called CopilotKit has released OpenDots. "OpenDots is completely open source and gives you the same thing [as OpenAI Dots] as a template you host yourself," the biz said in a post last week. "Bring any agent through AG-UI, use any OpenAI-compatible model, and make every part of it your own." The company previously released OpenMuse and OpenBot. These projects essentially separate the user interface, the AG-UI agent layer, and the backend to provide an alternative personal agent. This doesn't eliminate the token expense of whatever agent is deployed, but it does mean that OpenAI may not be doing the billing. Then there's a potentially lower cost option, called (confusingly) Open Dot, powered by AI models ru
Read full story →A phishing campaign targeting advertising managers by impersonating Gemini, Claude, ChatGPT, Perplexity, and Manus to steal credentials and multi-factor authentication (MFA) codes has added a fake Muse Ads product to its lure lineup – just eight days after Meta launched its personal AI agent. Meta announced Muse on September 8, and by September 17, a very convincing website – museads.ai – for a product called Muse Ads that promised to help advertisers reach buyers and run sponsored placements popped up online. “The operators already had the platform, so adapting it to a new brand can take minutes,” Oleg Zaytsev, lead security researcher at Island, told The Register. “The striking part is how quickly they turned a timely announcement into a credible reason for someone to act. The same platform could then be repackaged around other familiar tasks, from connecting a business tool to claiming a refund or applying for a job.” The security startup spotted the Muse Ads webpage, and upon digging into the scam uncovered that just the page was new. “Its code, from the sign-in forms to the fake browser window, came from a wider operation that had already run fake ad products for Gemini, Claude, ChatGPT, Perplexity, and Manus,” Zaytsev and fellow Island researcher Ofek Ronen wrote in a blog post published Tuesday. All of these products served as lures for browser-in-the-browner (BitB) attacks designed to trick agency staff, media buyers, and manager-account administrators into handing over their advertising account credentials – along with stored payment methods and client accounts – to digital thieves. “For victims, the potential cost is loss of access to an advertising account, unauthorized ad spend, and exposure of linked client accounts,” Zaytsev told The Register. How the scam works BitB is a clever phishing technique originally detailed by a researcher called mr.d0x in 2022. It involves building a fake login window directly inside a legitimate one. The fake window looks l
OpenAI is launching a new Intelligent UI feature in ChatGPT that allows the chatbot to answer your questions with interactive visuals. The update, which is rolling out to all users alongside GPT-6, gives ChatGPT the ability to combine a text response with diagrams, charts, forms, tappable buttons, and more. In a blog post explaining the […]
ChatGPT’s teen safeguards are meant to protect vulnerable users, but new testing found the chatbot continues encouraging engagement during crises and potentially encourages unhealthy relationships with the AI itself.
OpenAI is launching a new user interface that will bring interactive visuals to ChatGPT.
OpenAI’s Alexander Embiricos is coming to the AI Stage at TechCrunch Disrupt 2026, just days after the launch of Dots. Join this conversation by registering for your pass. Get you pass now to save up to $100 and get a second at 50% off.
Wrap Antigravity, ChatGPT Codex, Claude Code, Grok Build, Muse Code, Devin as an OpenAI/Gemini/Claude/Codex compatible API service, allowing you to enjoy the free Gemini Series, GPT Series, Grok Series, Claude model through API
At a high level, coding agents have a basic architecture: an LLM calls tools in a loop, while carrying forward the required context. Since that’s easiest to implement locally , many coding agents started out as terminal tools and then desktop apps (Claude Code famously began as a CLI tool). But as the open laptop meme illustrates, modern agent harnesses have frustrating limitations if they’re not able to be managed in the cloud . Since 2025, both OpenAI and Anthropic have been trying to shift their harnesses more into the cloud — but with varying levels of success. The challenges include keeping sessions reliable, isolating tool execution and preserving context. This sounds a lot like container orchestration in the early 2010s , from which emerged the open source Kubernetes in 2014. Kubernetes, developed by Google, became the dominant open source system for deploying and managing cloud applications at scale — it kickstarted the “cloud native” era of computing . Kubernetes uses control loops to keep applications running, recover from failures, and scale when needed. What if coding agents could be managed in the same way? That’s the bet that two Kubernetes creators, Craig McLuckie and Joe Beda , are making with Stacklok . Stacklok architecture; diagram via Stacklok Stacklok raised a $17.5 million Series A back in 2023 from Accel, Madrona and Bain Capital. Initially focused on software supply chain security , the company recently pivoted to Kubernetes-based agentic solutions . Mecatl: a harness built for the cloud The most interesting product of Stacklok is Mecatl, a “cloud-native harness” begun in June as an open source project on GitHub . At first, I thought the name was pronounced “me-cattle” — a play-on-words referencing the ‘cattle vs. pets’ concept of Kubernetes, where you treat your servers, pods and clusters like cattle rather than pets. But no, the name is actually pronounced “MEH-kah-tl” and is an Aztec word meaning a cord or rope ! In any case, what does clo
Every piece of AI content from Google's Gemini models has a hidden SynthID label, which makes it very difficult to pass the content off as authentic . SynthID has also gotten buy-in from other major AI players, but actually checking for those invisible watermarks has been a convoluted process. It's going to be a little easier going forward, though. Google has announced that its SynthID detector now supports watermarks from all its partners, and it's available for anyone to use on a new dedicated website. SynthID is an invisible signal encoded in the pixels of AI-generated images and videos, and in the waveform of AI audio. Google says Gemini alone has integrated SynthID with over 180 billion images and videos, plus 240,000 years of audio. Until now, the only way for those outside of Google's trusted tester program to check for SynthID was to ask Gemini. This is rather clunky, and the model's output could sometimes be confusing when you just wanted a simple "yes" or "no." Now you can cut Gemini out of the loop. Starting today, you can go to the new SynthID.com website to check a piece of content for an AI watermark. A major limitation of Google's previous detector was that it only checked for Google's version of SynthID. Even though ChatGPT images, for example, have SynthID, they wouldn't be flagged in Google's detector. That's no longer a problem. Read full article Comments
There’s been something of a change in the world of motorsport over the past few years. It has to do with where the money comes from. In the old days, it was tobacco and, to a lesser extent, booze. Today such things are mostly banned, and when you look at a racecar covered in sponsor logos, you’re likely to see something crypto-, cyber-, or AI-related. This makes sense—an ever-larger share of our GDP is wrapped up in those segments. They’re the ones with all the money, and it takes money to go racing. So at the amateur level, we’ve seen tech CEOs replace dentists as the privateer of choice: This past weekend’s Petit Le Mans included co-founders of GitHub, CrowdStrike, Shopify, and the creator of Ruby on Rails, to name a few. But this is an instance where a sponsor can provide tangible benefits to their motorsports partner beyond a bank transfer, as Chip Ganassi Racing (CGR) and OpenAI have been showing in IndyCar this year. Read full article Comments
Tickets for AIE NYC are selling out soon! See you next week! see past AINews issues for subscriber discounts. Pour one out for Mistral, who shipped a decent Large 4 “Le Chonk” model on the new 3800 GB300 cluster funded by their recent Series D . But they were overshadowed by more mathematics results from OpenAI’s internal Navier-Stokes math model - published as a blogpost , repo , and tweet . The best compliment comes from their Navier-Stokes competitor from Anthropic, who despite his personal issues with OpenAI, does not mince words: “It’s obviously the most significant moment in mathematical history.” This bears some qualification , but most experts seem to agree that it solves many of the top 500 open problems in math . In particular, Result 003, the Quasi-Riemann Hypothesis , is somewhere between a Fields Medal result and “the biggest result in number theory in 200 years ”. The most astonishing is the how - while Navier-Stokes was done in 88 hours and 10,000 agents , these solutions were 3 hours of ChatGPT Pro on average . AI News for 10/5/2026-10/6/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap OpenAI Releases 722 Math Manuscripts From an Unreleased Internal Model The release : OpenAI published a broad set of mathematical results from an internal frontier model in a public GitHub repo . It says it consulted the Institute for Advanced Study’s independent Advisory Group on Mathematics and AI on how to release them. Scale and compute : The collection reportedly holds 722 manuscripts grouped into 372 families of related results. They came from an evaluation of about 4,000 research problems and used an average of roughly three hours of ChatGPT Pro thinking compute per result ( summary , Rundown ). Artifacts : The release includes papers, proof artifacts and selected reasoning summar
OpenAI has revealed solutions to a number of long-standing mathematics problems produced by an unreleased frontier model in a batch of 722 manuscripts, covering 372 result families that group related papers. It extends a run of breakthroughs that have both impressed and unsettled parts of the mathematical community while raising questions about research ethics and […]
OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union, the company has announced. It will also offer the watermarking feature in other regions, but it will be off by default outside of the EU. The move in Europe is driven by a need to comply with the EU AI Act, which took effect in August. It requires marking content produced by AI models in a way that another tool can detect. Unfortunately, there is still no completely effective and reliable way to do that. A few standards already exist, like SynthID and the C2PA project, but they are relatively easy to circumvent for anyone with basic know-how. The same is likely true for OpenAI's watermark. Its method is proprietary; the company calls it textGrain, and has published a technical paper explaining how it works. But in general, it works like other LLM watermarking tools we've seen in the past: It puts patterns in the word choices that are not clear to a human reader, and that don't meaningfully change the general quality of the output, but that someone with a key can use a specialized detector to find. OpenAI says it will be giving access to the detector to a limited number of researchers and organizations, and providing a request-for-approval process for others to be added over time. Read full article Comments
An independent briefing for builders: the whole field read continuously, every story scored for relevance, and the noise left off the page.
300+ curated sources. Every story scored 1–10 for builder relevance by Claude's frontier model. The filler never makes it to the page.
GPUs, datacenters, power deals, and inference economics: the infrastructure layer that decides what every builder pays. Our signature coverage.
Every story is sourced. Every score is computed. We show our work and link to originals.
Every briefing closes with The Call: one falsifiable claim with a date on it. When we're wrong, we say so in print. Opinions are cheap; ours get scored.