OpenAI DevDay 2026 ran on Tuesday, September 29, at Fort Mason in San Francisco. The keynote started at 10:00 PT, which made it an evening event here in Istanbul (20:00). OpenAI counted more than 20 major announcements and said ChatGPT now has 1.2 billion weekly users. The headliners were dots (always-on agents on GPT-6 Astra), GPT-6.1 Sol, a speed tier called Ultrafast and a $500 Pro plan.
My short version: dots got the stage, but the story is the meter. Nearly every launch that matters to a developer is about who pays for tokens, how much and through which account.
Some context before the product list. A day earlier OpenAI said it would not ship GPT-6.1 Astra, and I wrote about that decision and the safety-case rules behind it in a separate deep dive. I won’t repeat it here. It does matter for how you read dots, though.
📋 What OpenAI DevDay 2026 actually shipped
Availability varied a lot, so here is the list with the fine print attached:
- Dots: rolling out to Pro users (markets outside the EEA, Switzerland and the UK) and Business Premium. Enterprise, Edu and Healthcare get a beta that admins must switch on.
- GPT-6.1 Sol: live in the API as
gpt-6.1-solat $2 input, $0.10 cached input and $10 output per million tokens. In ChatGPT it lives in Work and Codex only, “not yet available in Chat.” - Ultrafast: GPT-6 Astra at up to 300 tokens per second. $60 in and $300 out per million tokens in the API. Sol Ultrafast is “coming soon.”
- Pro 500: $500 a month, 25 times the Plus allowance, Ultrafast included.
- Agents API with computer use: the Agents API has been in public beta since September 10. DevDay added computer use.
- Decisions API: limited preview, powered by GPT-6 Luna.
- Sign in with ChatGPT: identity for everyone, plan usage for Plus and Pro users in 16 partner tools.
- Also: Codex in the cloud, a refreshed Codex CLI with voice, Code Review in the desktop app, Codex Security Cloud, Bedrock Managed Agents, plugin extensions, MCP Events, ChatGPT Space, Pages, Private Intelligence and an enterprise Marketplace with 32 partners.
That is a lot for a 50-minute keynote. Some of it is real and usable today. Some of it is a waitlist wearing a launch badge.
🫧 Dots: good design, confused pitch
Dots are OpenAI’s answer to Meta’s Muse, SpaceXAI’s Grok Bot and Instinct. Each dot runs on GPT-6 Astra, has its own cloud computer and browser, reaches 4,000+ apps through plugins and keeps context across ChatGPT, Slack and Teams. You can call it by voice. Texting is “coming soon.”
The detail I care about: it is GPT-6 Astra, the model that already shipped. Not the 6.1 version OpenAI held back on Monday. Sam Altman called Astra “our most aligned model” on stage, and after the week OpenAI just had, that sentence was doing a lot of work.
The safety design is better than I expected. According to OpenAI’s dots safety post, background “proactive research” runs with read-only tools. Actions that touch your accounts pass through an auto-review against your instructions and Custom Rules. A monitor can pause or stop a dot. Access to your own laptop starts switched off, and a password change always stays with you. If you run agents in production, that is roughly the checklist you would write yourself.
The pitch is the problem. Benedict Evans, quoted in Platformer, called the launch video “rather confused”: “childish anthropomorphism, hardcore startup software-engineer use-cases randomly mixed with wedding planning.” I agree with him. The Hacker News thread (571 points, 426 comments) opened in the same key: “I read it a few times and still have zero idea what this thing is.”
And the live demo stalled. Simon Willison’s live blog caught it: “Dottie is having a slow morning,” followed by an awkward silence. Casey Newton, after a couple of hours of access, wrote that dots impressed him more than the other agents he had tried. Both things can be true.
Why Pro only, when Muse is free? My guess is cost. An always-on agent running a frontier model eats compute around the clock. CFO Sarah Friar told CNBC the vision is to bring dots to the whole consumer base. Not yet, though.
💸 GPT-6.1 Sol: the price cut is mostly a cache cut
GPT-6 Sol came out on September 22. GPT-6.1 Sol replaced it seven days later.
OpenAI’s headline is “near-Astra intelligence at a fifth of the price.” Look at the pricing page and the fifth is measured against Astra’s $10/$50. Against GPT-6 Sol, the list price hasn’t moved: both are $2 in and $10 out. The one change is cached input, which drops from $0.20 to $0.10 per million tokens.
That sounds like a footnote. For agents it isn’t. An agent loop resends the same context again and again, so the cached rate ends up running your bill. Hacker News user minimaxir put it best: “This is the actual big announcement. 50% cheaper cache than GPT-6 Sol will get you far more mileage on Codex.”
The capability jump looks real on paper. Artificial Analysis puts 6.1 Sol one point below Astra on its Intelligence Index, at $0.72 per task against Astra’s $3.26 (max effort). It also found 6.1 Sol uses 10 to 30 percent more output tokens than 6 Sol, so check your own bills. Astra still tops OpenAI’s own Terminal-Bench Science chart at 68.1 percent.
My problem is the churn. A model that lives for a week breaks eval baselines, prompt tuning and cost forecasts. One HN commenter wrote, “GPT-6 Sol released a week ago. Shortest model life ever?” If you built on 6 Sol last Tuesday, you now have homework. I covered the Sol and Luna cost curve last week, and this price sheet bends it further down for cached workloads while leaving everything else where it was.
⚡ Ultrafast and Pro 500: speed is now a line item
Ultrafast runs GPT-6 Astra at up to 300 tokens per second, up to 8x faster in Codex and 6x in the API. It costs six times the standard rate. On stage, Altman’s defense was “you know what, it’s worth it.” Thibault Sottiaux later said the team used Ultrafast to help merge the ChatGPT and Codex desktop apps in 28 days.
A detail for anyone in Europe: the Ultrafast docs say it supports US data residency and global processing only. No EU regional endpoint.
Then the plans. Pro 500 costs $500 a month and gives 25 times the Plus allowance plus Ultrafast. At the same time, Pro 200 reopened to new subscribers with a lower allowance. Business Insider reports it drops from 20x Plus to 10x, and GPT-6 Pro messages go from 200 to 100 a week. Eligible existing subscribers keep the old allowance until October 29, per OpenAI’s help page.
Sottiaux did warn people the night before on X: the new math “will net out at half the dollar in API spend compared to the old Pro $200 plan.” I give him credit for saying it plainly, before the keynote. HN still called it a rug pull, and the arithmetic backs them up. One commenter summed it up as “20x” becoming “10x,” and “25x” for two and a half times the money.
My read: OpenAI is done subsidizing heavy Codex users at $200 and is pricing on outcomes. That’s defensible. What isn’t defensible is selling a $500 plan without publishing what “25 times Plus” means in tasks or tokens. As one commenter put it, “They still refuse to tell you exactly how much usage you’re paying for.”
🔑 The quiet launches that matter more
Sign in with ChatGPT is the one I’d circle in red. Plus and Pro users can spend their plan allowance inside 16 partner tools, including Cognition’s Devin, Notion, Vercel, T3, OpenClaw and Dactyl, with a cap per app. According to the help page, identity sign-in shares only name, email and avatar, and spending your plan is a separate consent.
For indie apps this is big. You can ship an AI feature without charging users for tokens a second time. Willison wrote “I’ve wanted this one for years!” and Altman said in the closing Q&A that “we should have done it a long time ago.” The flip side: if your margin comes from reselling tokens, OpenAI just put a toll booth in front of your checkout.
Decisions API hands Luna a fixed set of answers and asks it to pick one, for classification, routing or an agent’s next step. The New Stack quotes around 150 ms, against 1.6 seconds for plain GPT-6 Luna. There’s no public price yet. Willison said it “sounds like their response to Jev,” TypeSafe’s decision model that I’ve been following closely. When OpenAI answers a new category within two weeks of its debut, the category is real.
Codex Security Cloud had the best numbers of the day, from a breakout session: an internal sprint fixed 53 critical findings on day one, 36 percent of findings were duplicates, and generated patches had about a 1 percent rollback rate. Bedrock Managed Agents puts OpenAI’s agent harness inside AWS, which will settle a lot of procurement arguments.
🏆 Winners and losers
Winners:
- Indie builders who adopt Sign in with ChatGPT early.
- Teams running cached agent loops on Sol.
- AWS, which gets OpenAI agents running natively in Bedrock.
- The decision-model category, now that OpenAI has validated it.
Losers:
- Pro 200 power users after October 29.
- Anyone in the EEA, Switzerland or the UK who wanted dots on Pro.
- Apps whose business is marking up tokens.
- Teams with slow eval pipelines, because a one-week model life punishes them first.
The in-between case is Anthropic. One HN commenter felt “the Opus 5.5 mog cast a pall on the room,” and OpenAI’s Sol post names Opus 5.5 on GDP.pdf, AutomationBench and cost per science task. That tells you who OpenAI thinks it is racing.
🧭 What HN and X got right, and what they got wrong
HN got the money right. The most useful comments were about cache pricing and plan allowances, not dots mascots. They also spotted Sign in with ChatGPT as the sleeper: “Subscription sharing is going to be huge for indie apps that don’t want to deal with token billing.”
The biggest mistake was mixing up the models. An early comment on the Sol thread asked, “Weren’t there headlines just yesterday that they weren’t releasing this due to safety concerns?” No. The held model is GPT-6.1 Astra. GPT-6.1 Sol is a different, smaller model, and dots run on GPT-6 Astra. The naming makes this easy to get wrong. That’s on OpenAI.
HN also went too far with “none of their 20 new features seem genuinely usable.” Codex Security Cloud, the Agents API and Sign in with ChatGPT are usable today. And the theory that 6 Sol was a renamed smaller model is speculation, so I’m leaving it there.
🛠️ What I’d do this week
- Re-run your evals on
gpt-6.1-soland compare cost per task, not the rate card. - Measure what share of your input tokens hit the cache. That number decides your savings.
- If your app charges for AI usage, read the Sign in with ChatGPT docs before your next pricing decision.
- Test Ultrafast only on paths where a user is actually waiting, and budget it as premium.
- Before you give a dot access to anything, write your own Custom Rules first.
Altman told reporters OpenAI won’t go public until it can “make confident safety claims,” per The Verge. On the same day it shipped an always-on agent and a pricier plan. Those two things sit side by side at OpenAI now. For builders, the practical answer hasn’t changed: take what’s live, price it yourself, and ignore the mascots.
📚 Sources
- OpenAI, DevDay 2026 Recap
- OpenAI, Introducing dots and dots safety
- OpenAI, Introducing GPT-6.1 Sol and API pricing
- OpenAI, Ultrafast mode docs and Pro tiers help
- Simon Willison, DevDay 2026 live blog
- Hacker News: GPT-6.1 Sol, Dots, Pro 500, Recap
- Platformer, The Verge, Business Insider, The New Stack, Artificial Analysis
More context: AI weekly, 21-27 September.
1 comment