Dots, Gemini 4 Argon and Sonnet 5.5: What Happened in AI's DevDay Week

The chatbot era is clocking out and the agent era is clocking in. That single sentence sums up the week; the rest is not trivial, but telling which detail matters takes work. OpenAI made more than twenty announcements at its developer day, Google shipped a new flagship after months of silence, and Anthropic quietly cut prices. According to the Axios rundown of DevDay, the real push on stage is the shift from software that answers questions to software that takes on jobs, and that shift has moved from marketing slogan to product list.

First, the Google front. Gemini 4 Argon is a frontier model built for long-horizon knowledge work such as legal, finance, and cyber defense. In the tables the company shared, it leads GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across most knowledge-work tests. The encouraging part comes with a sobering one: the model is open only to selected cyber defenders and trusted testers for now. As Google explains on its blog, the phased rollout reflects guardrail work and a voluntary pre-release process with the government. Pricing is set at 2 dollars per million input tokens and 10 dollars for output; in-house results are listed too, from beating a published quantum baseline by 40 percent to fleet-wide memory savings above 300 tebibytes. The speaker's reservation is fair: by the time the door opens to everyone, rivals may already be one model ahead.

The mood going into DevDay was sour. GPT-6.1 Astra was pulled at the last minute over safety concerns and pushed to October, and the 200-dollar Pro plan saw its 20x usage promise cut to 10x. Wounds from the previous two models were fresh: GPT-6 Sol had disappointed, while GPT-6 Astra burned through daily usage too fast. As Axios reports, the delay landed in the middle of the debate over whether increasingly autonomous systems can be audited. In other words, the team walking on stage owed an apology rather than expecting applause, and the tone of the presentation showed it.

The headline act was Dots. In OpenAI's own words, this is an always-on helper: it learns from feedback, acts proactively, owns its cloud computer and browser, and plugs into more than 4,000 apps. It draws its power from Astra and currently runs inside the Pro plan without drawing down usage. The speaker's demo shows Dots living as one long conversation inside the chat app; it can be renamed, restyled, and watched while it works on its computer. It starts singular and is meant to grow into teams of dots over time. It reads as a direct answer to Grok Bot and Meta Muse.

So what did Dots do in 24 hours of real use? The speaker had it scan Gmail and calendar, dig up forgotten tasks, and enjoyed the proactivity of unprompted check-ins through the day. But the experience is rough: messages sometimes go unanswered, replies sometimes arrive late; the product clearly shipped in a hurry. The most interesting finding is positioning: Dots works less like a personal aide and more like an orchestrator . It delegates coding jobs to Codex and stands guard over them; the user trusts the watch instead of chasing every task. Grok Bot still leads at setting up automations that run around the clock, Muse still leads at personal accompaniment. Access stays limited to Pro and Business Premium, and the free-usage promise covers only the first month; what comes after is unwritten. That is the fine print on the OpenAI page: nobody wrote down what the free first month costs in the second.

The second big reveal is ChatGPT Space, a shared workspace where teams, agents, and dots meet on the same document. Pages, the new document type, are visually rich and templated, able to generate tables and images with generative AI; at first glance they feel like Notion. A meetings add-on records calls, produces summaries with action lists, and promises to delete the audio once done. Images and sites previously made inside chat gather in the same area. The speaker doubted its purpose at first sight, then found scenarios like project tracking after poking around. As described across the OpenAI help pages, Space sits somewhere between Slack and Drive, with agents embedded throughout.

There was relief on the model front: GPT-6.1 Sol arrived. One name correction comes first, because the video says Soul while the real name is Sol, and the DataCamp review confirms it. Sol is the fast revision of GPT-6 Sol, which had disappointed a week earlier, and it delivers the promise: near-Astra intelligence at roughly one-fifth of the price. The figures are on the table: 2 dollars per million input tokens, 10 dollars for output, a context window above one million, and a Terminal-Bench Science bill cut from 23.80 dollars with Astra to 5.47 dollars. It is open to Plus subscribers inside Codex and Work; the speaker made it his new default. The launch-day slowness was attributed to crowding and now runs faster. One fine detail: Sol does not support the lowest reasoning settings, so anyone moving over from GPT-6 Sol may need small code changes.

The pricing tier, though, feels harsh. The new 500-dollar monthly plan brings the highest limits with an Ultrafast mode that promises snappier Astra answers across Work and Codex. According to The Verge, the 200-dollar tier was reopened with new limits. The speaker's frustration is understandable: usage trimmed from the 200-dollar plan looks resold for money in the 500-dollar one, and Ultrafast burns through limits fast. He chose to stay on the 100-dollar plan and rely on the frugality of Sol. The same package carries two practical extras: signing into third-party apps with a ChatGPT account and sharing limits, so tools like the Notion AI add-on run off the chat quota without a separate fee, and Codex moving to the cloud, so a phone can drive work running elsewhere.

The speaker's three-part verdict on DevDay is sharp. First, the event was overhyped; there was no single breakthrough, only good updates. Second, expectations were already bruised by the Astra delay and the trimmed plan. Third, the differences between the new tools were poorly explained; the line between Dots, chat, shared work, and Codex stayed blurry. Peel off those three layers and three solid things remain: a promising Dots, an intriguing Space, and a genuinely good Sol model. That habit of filtering hype is worth keeping in your pocket for every launch week.

The quiet star on the Anthropic front was Sonnet 5.5. The company announcement frames it as a clear step over Sonnet 5, with 30 percent faster output and up to 30 percent lower cost per task; the price tag is unchanged, savings come from spending fewer tokens. The numbers stand out: 70.6 percent on Terminal-Bench 4.0 against 10.3 for Sonnet 5, and two points behind Opus 5.5 on a real-world occupational test. It is also the first Sonnet to finish a game working from screenshots alone, backing the long-horizon claim. The speaker's split is neat: Opus for strategic planning, Sonnet for execution. For cheaper plans it could be the week-saving default. Opus for costly work, Sonnet for daily work, and Haiku 5.5 coming for high volume; the trio in the Anthropic note reads tidy against rivals' scattered price lists.

Google shipped two small updates that change daily life. First, skills in Gemini: the most repeated instructions are saved once and called with a slash, stacked together, and enriched with reference files such as text and PDF documents. Skills will gradually replace Gems, with existing data migrating over. Second, NotebookLM becomes Gemini Notebook with a visual refresh: more expressive voices in audio overviews thanks to new models, promises of more languages and longer non-English overviews, and voice chat with notebooks rolling out fully to Pro users on mobile. As the Google skills guide puts it, the point is to stop rewriting the same request every day; the productivity promise is as real as paperwork.

Two stories on the Meta front, one cheerful and one cautionary. The cheerful one: Muse opens to small business. As TechCrunch reports, the agent plugs into Shopify, Dropbox, Slack, Notion, Stripe, Canva, and QuickBooks, helping with sales and marketing by knowing what the business sells, how the brand sounds, and what customers ask; limited use is free, heavier use is subscription. The cautionary one is the viral Marketplace episode: after the user left permissions on always allow, the assistant accepted a lowball offer, shared the home address, and booked a pickup; the wrong figure was later explained as a typo. Three lessons in short: never open broad permissions blindly, never hand the assistant data like an address, always keep the final approval on sales. Companies need permission designs that prevent such accidents before they happen.

The closing act brings a surprise name: Q from Madness. Announced a day or two before Dots, Q runs on a Grok Bot-like line but pushes further: alongside its own computer, each bot gets its own email, phone number, and wallet. Tasks can be assigned by email, it can hold the line on phone calls, and it can shop within limits on single-use cards; it runs a little more automatically while the user keeps the final approval, and multiple bots can gather in a group chat. An honesty note: the ambitious Q features could not be verified against independent sources beyond the video narration and launch material; single-announcement novelties count as unverified. The speaker's favorite of the week is still Dots; read together with the orchestrator role, the new models, and Space, the week's headline is clear: AI no longer answers, it takes on work. Which is why the question changed: which job, to whom, and with what permission?