Yuval Avidani
Author
Yesterday, September 29th, OpenAI took the stage at DevDay 2026 in San Francisco with more than 20 announcements - and its two biggest ones already exist on the market. dots agents arrived seven weeks after SpaceXAI's Grok Bot, and the Decisions API, a model that makes decisions without writing text, showed up two weeks after Jev from the startup TypeSafe. And on the side, almost without any noise, the $200 subscription got half its quota cut.
Let's break this down: what actually launched, what we've already seen elsewhere, and how much more this is costing us now compared to a week ago.
What launched yesterday, in short
Before we dive in, here's the full list of what matters:
- dots agents - always-on agents with their own cloud computer, running on GPT-6 Astra.
- GPT-6.1 Sol model - just one week after GPT-6 Sol, better at coding and computer-use, at the exact same price: $2 per million input tokens and $10 for output.
- Decisions API - a model that returns a decision instead of text, in a limited preview version.
- Ultrafast mode - 6x higher speed via the API, at 6x the price.
- Pro 500 subscription - $500 a month, alongside a cut to the $200 tier.
- And more: a managed Agents API, plugins that are basically full apps inside ChatGPT, a preview of Private Intelligence, and ChatGPT Space for collaborative document editing.
dots agents: the assistant that never goes home
Let's start with the definition. A dot agent is an always-on AI agent that gets its own cloud computer (Linux with a Chrome browser), works in the background toward goals you gave it, and talks to you through ChatGPT, Codex, Slack, Teams, and even a phone call (SMS is coming) - for Pro and Business Premium subscribers. Think of it like a personal assistant who has their own desk in the office: you don't ask a question and wait around, you hand off a task and they come back to you only when they need approval.
In the on-stage demos, a dot named Alfred finished a phone call with the words "Done, you're live" - the user's website had gone live. Another dot edited a dashboard by voice, and along the way noticed on its own that a finance meeting clashed with the user's daughter's recital, and suggested moving it.
Sound familiar? Because it is. Almost every core feature in dots already existed in Grok Bot, which launched in beta on August 11th: a persistent cloud computer with a browser, agents with names and personalities, Teams integration, approval rules where you decide what the agent can do on its own versus what needs your sign-off, and even the promise that using the agent wouldn't eat into your regular subscription quota. The product cadence, the marketing language, the order features shipped in - it all lines up.
To be fair, there are real differences too. The first is security: Grok Bot's documentation explicitly states that all bots belonging to the same user share one computer and "are not security boundaries." At OpenAI, every dot gets its own computer - looks like they learned from a weakness the competitor disclosed about itself. The second is phone calls, which we haven't seen in Grok Bot yet. And the third is context: both Meta with Muse and the open-source project OpenClaw were there before everyone else, so the idea of "an agent that lives in the cloud" doesn't belong to any single company. What is clear is that OpenAI wasn't leading here - it was the last of the big players to arrive.
And now for the part where the comparison stops flattering OpenAI - the price:
You can get Grok Bot starting at $20 a month through Cursor Pro. You can only get your first dot with a Pro subscription, which starts at $100. And a few more caveats worth knowing: every additional dot will cost money, at a price that hasn't been published yet. You can't view or delete a single memory the agent saved - only reset the entire dot. And on Pro subscriptions, it's not available in the European Economic Area, the UK, or Switzerland. We couldn't find an official mention of Israel, so it's worth checking before you upgrade.
The Decisions API: a model that doesn't write a single word
This is where it gets more interesting, because we're talking about an idea most people don't know yet. Turns out a big chunk of what AI does inside applications isn't writing at all - it's small decisions: which queue to route a request to, whether a message is spam, how urgent a request is. Today we ask a language model to write an answer, and then extract the decision out of it. Think of it like asking a novelist to write a whole essay just so you can check a box on a form - slow, expensive, and sometimes he writes something your code can't even parse.
The solution: a decision model. You define a question and a closed list of possible answers, and it returns one answer with a probability - without generating any text at all.
And here's where Jev's story comes in. Jev is a decision model from the startup TypeSafe, which launched in early access on September 15th - and it doesn't produce a single token. One of the founders is Diogo Almeida, a former OpenAI researcher who worked on ChatGPT and RLHF. Instead of guessing the next word over and over like a regular language model, it reads the entire input and returns a decision directly, in three forms: Choice (pick from a list), Score (a rating on a scale), and Noul (probability of "yes"). The price: $0.042 per million input tokens, and output is free. Its limitations are also written openly: it struggles with precise arithmetic, counting, date comparisons, and irrelevant context, and it only accepts text. And Check Point already showed it can be attacked with prompt injection (a malicious instruction hidden inside the input), exactly like a language model - with up to a 93% success rate in the best attack.
Two weeks later, OpenAI unveiled the Decisions API: you define a question with a closed set of answers, send context as text or image, and get back a decision. According to OpenAI, it responds in about 150 milliseconds, ten times faster than GPT-6 Luna through the regular API (roughly 1.6 seconds).
So is it a copy? The idea, the timing, and even the name all point that way. But it's worth looking under the hood: according to reports, the Decisions API runs on a version of GPT-6 Luna, meaning a language model optimized for decisions, not a new architecture like Jev's. That's the difference between building a new model from scratch and taking someone else's idea and wrapping it around the model you already had. To OpenAI's credit, it also supports images, which Jev doesn't. And what's missing: pricing. OpenAI hasn't published how much the Decisions API costs, so for now there's no way to know if it even competes with Jev's $0.042.
Subscriptions: what was and what will be
And now, to our money. Before DevDay, OpenAI had three consumer subscriptions that include Codex and Work (there's also Go at $8, without those tools): Plus at $20, Pro at $100 with 5x the usage of Plus, and Pro at $200 with 20x. On September 10th OpenAI paused new sign-ups for the $200 tier, saying it "put the most strain on our systems." Now it's reopened - but different:
- Pro 200 subscription - drops from 20x to 10x Plus usage on Codex and Work, and GPT-6 Pro messages drop from 200 to 100 a week. Anyone signing up now gets the new quota immediately; existing subscribers keep the 20x rate until October 29th, transition on October 30th, and get a one-time credit valid through the end of the year.
- Pro 500 subscription - new: 25x Plus, and the only personal tier with Ultrafast (up to 300 tokens per second on Codex, 8x the normal rate).
- Pro 100 subscription - stays at 5x.
- All Pro tiers - now include one dot, and the 5-hour window on Codex and Work has been removed, so you can use your weekly quota whenever you want.
Let's do the simple math: how much does one "Plus unit" of usage cost on each tier. This is the number that actually shows what changed:
Until now, whoever bought the $200 tier got every usage unit at half price - $10 instead of $20. After the change, every personal OpenAI subscription that includes Codex and Work costs exactly $20 per unit, and the bulk discount disappears. In other words, Pro 500 isn't a "great upgrade" - it costs 2.5x Pro 200 and gives you 2.5x the usage. What you get on top of that is speed (Ultrafast), not value.
On the developer side, the picture is similar: GPT-6.1 Sol improves at the same price, and that's genuinely good news. But Ultrafast costs 6x more, so GPT-6 Astra jumps from $10 and $50 to $60 and $300 per million tokens. Feel free to run the numbers yourself:
Bottom line
In my view, DevDay 2026 wasn't an event about invention - it was about distribution and pricing. OpenAI wasn't first with an always-on agent, and it wasn't first with a model that decides without writing. But it knows how to take an idea that proved itself with a smaller player and serve it up to ChatGPT's massive audience. That's a real skill - it just deserves to be called by its actual name.
The genuine innovation here is really in the price engineering: the tier that gave the most value per dollar got cut, and anyone who needs more now has to jump to a $500 tier. And remember the caveats: the Decisions API's price hasn't been published, the price of an extra dot hasn't been published, and availability in Israel hasn't been officially confirmed.
So if Grok Bot gives you an agent for $20 and Jev gives you decisions for 4 cents per million tokens - what exactly are we willing to pay OpenAI for, beyond the name?