OpenAI's Biggest DevDay was Underwhelming, and Google's CEO is a "Monster"

OpenAI’s headline DevDay announcement was Dots, an always-on personal agent powered by GPT-6 Astra. Each Dot gets its own cloud computer and browser, so it can keep working while your laptop is closed. OpenAI says its plugin ecosystem connects to over 4,000 apps. Give it an ongoing responsibility, such as tracking a project or preparing work for review, and it follows through between conversations. OpenAI’s announcement

That puts OpenAI directly against Meta’s Muse and Grok Bot. The useful promise is continuity: less time explaining the same project, more work ready when you return.

The live demo had trouble, including a delayed response when presenter Holly Li tried speaking to her Dot, “dottie.” An assistant that works 24/7 still managed to be unavailable for the few seconds its employer needed it most. Axios’s event coverage

The model announcement was GPT-6.1 Sol. OpenAI says it approaches Astra on coding, computer use, and professional work at one-fifth of Astra’s standard token prices. Its tests also put Sol ahead of Opus 5.5 on some document and business workflow tasks. At $2 per million input tokens and $10 per million output tokens, that could make longer, repeated tasks much cheaper. Model announcement

The missing flagship was GPT-6.1 Astra, which OpenAI held back following safety testing. CBS News DevDay delivered plenty of products, but the frontier model jump some people expected will have to wait. Personal agents look likely to stick around. The model rankings and prices could change before you finish choosing a subscription.

OpenAI introduced Pro 500 at $500 a month, with its highest included usage and access to Astra Ultrafast. The $200 plan is available to new subscribers again, with a lower allowance for those without grandfathering. Official plan details

Speed has a cost beyond the subscription: Astra Ultrafast consumes included usage at eight times the Standard rate. Cheaper everyday models and expensive premium access are becoming part of the same business. For heavy users, the amount of work a plan covers matters as much as its monthly price. Usage and speed billing

Opus 5.5 remains the strongest option for demanding work. Anthropic’s launch tests showed gains in coding and knowledge tasks, with typical workload costs around 40% below Opus 5. The new Sol results make a universal “Claude wins everything” claim difficult to defend, but Anthropic’s combination of capability and efficiency remains much more appealing imo. Opus announcement

Sonnet 5.5 arrived on September 28. Anthropic says it generates output over 30% faster and costs up to 30% less per task than Sonnet 5. Its token prices stayed the same; the savings come from using fewer tokens. It also lands close to Opus on Anthropic’s professional work evaluation, giving users a promising cheaper option for routine documents and coding tasks. OpenAI had the big event. Anthropic still found a way into the conversation. Sonnet announcement

AMD has agreed to buy Fei-Fei Li’s World Labs for approximately $8.2 billion in stock. The deal is expected to close by the end of 2026, subject to approvals, after which Li would become AMD’s executive vice president and chief scientist. World Labs builds models that generate and simulate interactive 3D environments for applications such as robotics. The deal gives AMD research expertise to help design hardware for emerging AI workloads, extending its competition with Nvidia beyond chips alone. AMD’s announcement

Meta announced Muse for Small Business, adding connections to services including Shopify and QuickBooks. It can use business information to prepare plans and handle tasks, while publishing, sending, and spending require approval. Muse is available in the US and Canada. Meta has progressed from supplying Facebook boomers with AI slop to offering help running their businesses. This probably killed another 67 adtech startups before their founders finished the launch thread. Meta’s announcement

Smaller companies are pursuing the same opportunity. Instinct lets people request tasks through iMessage or WhatsApp, making personal agents accessible through an interface they already use. The Atlantic’s reporting

Nvidia’s Open Agent Safety Platform combines OpenShell, software that enforces boundaries around agent activity, with Sentry, a separate watchdog running on BlueField-4 hardware. Nvidia says it can quarantine agents that cross those boundaries in milliseconds. Companies giving agents access to important systems need controls the agents cannot change. As everyone sells assistants that act independently, selling the emergency brake looks like a sensible business too. Nvidia’s announcement

Z.ai’s GLM-5.3 is attracting attention for its coding and cybersecurity abilities. NIST’s CAISI called it the most capable open-weight model for cybersecurity released at the time of its assessment, estimating a gap of around four months behind the US frontier on its cyber benchmarks. NIST’s assessment

Anthropic’s research found its safeguards easy to bypass: under different test conditions, the model engaged with harmful requests in 64% to 100% of attempts. Those figures measure willingness to proceed, rather than successful hacks. Anthropic’s research For businesses, the appeal is lower cost and more control over deployment. GLM-5.3’s listed API prices are $1.40 input and $4.40 output per million tokens. Open weights offer more hosting options, although licences and operating costs still need checking. Z.ai pricing

Visiting dot.com redirects to Grok Bot. xAI owns it, with a July transfer; a deliberate prank remains unconfirmed. OpenAI spent DevDay explaining its new agent while the matching singular domain sent visitors to a rival. Someone had a much cheaper launch campaign. Domain reporting

David Pogue’s 125 tests show a more useful Siri, capable of retrieving personal information and acting on content on screen. Finding a meeting link or pulling information from an email could save professionals plenty of daily fumbling. His examples also show limits, so there is still room between a good demo and an assistant you trust with everything. Pogue’s tests

Meanwhile, Apple’s iOS 27.0.1 update fixes iPhone 18 Pro restarts after failed Face ID authentication, colour artefacts in certain photos, and a touchscreen issue involving the notification and control panels. A better Siri is welcome. A phone that survives unlocking is welcome too. Apple’s release notes

Trump’s September 29 meeting with tech executives produced a voluntary AI safety accord covering internal reviews and outside audits. He described it as “morally binding,” with more details still to come. Trump also continued pushing the name “Super Intelligence,” which Jensen Huang subsequently used at a panel. Also, Sundar Pichai was labelled a “monster”, with viewers given a free hand to decide on its meaning. Accord reporting

When a reporter asked whether self-policing was enough, Trump interrupted with “fake news,” while Zuckerberg and others laughed. Amodei had just acknowledged real risks and said the response remained under discussion. A question about holding the industry accountable became a group laugh with the industry standing behind him. Reporting and video

InstaCloud handles the part after a coding agent writes your app. It lets agents manage cloud infrastructure through a command line, scales compute down to zero when idle, and clones environments so changes can be tested away from production. Founder Hang Huang announced an $8 million seed round this week. Product, funding reporting

The free tier includes $10 in monthly usage credit. For anyone building small tools with coding agents, it offers a place to get an app online and let the agent handle more of its upkeep. “The code is ready” has never guaranteed the end of your evening. Free tier details

That’s it for this week. See you next Wednesday.

Saksham