On Monday 28 September OpenAI apologised to Australia. One of its agents had opened files on a Medicare statistics portal in June, and the company said, "We are sorry and working to do better in the future." The same day it withdrew GPT-6.1 Astra, a model it had planned to release, because it "did not meet its safety standards."
On Tuesday it held DevDay at Fort Mason in San Francisco, in front of about 2,500 people, and the headline launch was an agent that runs all the time. OpenAI calls them dots.
The recap lists 25 launches. I read it, the product posts, the pricing pages, the GPT-6.1 Sol system card, the help articles and the keynote's captions. Four of the 25 matter most if you build on OpenAI or buy it for a team, and I go through those in detail: dots, GPT-6.1 Sol, the Ultrafast speed tier, and the new plan prices. The rest are in a table further down.
Dots
A dot runs on GPT-6 Astra, OpenAI's top model. Each one gets its own cloud computer, connects to more than 4,000 apps through plugins, and, in OpenAI's words, "can work towards your goals 24/7." You reach it in ChatGPT, Slack, Teams or by voice call. Texting is a beta for Pro users in the US. At launch a dot cannot have its own email address and cannot call you.
Dots at launch | |
|---|---|
Model | GPT-6 Astra |
Who gets one | Pro users (the $100, $200 and $500 plans) and Business Premium seats |
Where | Everywhere ChatGPT is sold except the European Economic Area, Switzerland and the UK |
Enterprise, Edu, Healthcare | A beta, off by default, that a workspace admin has to turn on |
Price | The first dot is included in the plan. More dots, and more speed per dot, are promised "in the future" |
Usage | Talking to your dot does not count against plan limits. Codex and ChatGPT Work tasks the dot starts do |
Altman's pitch was about attention: "I feel like I've finally gotten some of my attention back. I no longer feel quite as addicted to my phone in the same way." Alexander Embiricos, on OpenAI's product staff, described where the dot stops: "We're not going to do this for you, but we can take you all the way up until the point where you do it yourself, and we'll hand over to you." Fortune reports that dots ask permission before sensitive actions.
Meta launched Muse, which does much the same thing, on 8 September, with a free tier. SpaceXAI's Grok Bot has been in beta since 11 August. Dots start on plans that cost $100 a month or more. Benedict Evans called the launch "a rather confused product launch," mixing consumer branding with "hardcore startup software-engineer use-cases." TechCrunch pointed out that most of what a dot does was already possible in Codex, and that dots bundle it into one package.
The demos did not help. Simon Willison, live-blogging from the room, wrote that "there were a whole lot of demo failures during DevDay," including a dot that sat silent on stage and a voice demo that failed.
GPT-6.1 Sol
OpenAI describes GPT-6.1 Sol as an upgrade to GPT-6 Sol that "nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices." It is available in ChatGPT Work and Codex on paid plans, and in the API as gpt-6.1-sol. It is not in regular ChatGPT chat yet.
On DeepSWE, a software-engineering benchmark, 6.1 Sol's best score is 75.2% at $0.65 a task, against 74.1% for Astra at $4.43. More effort makes it worse after "high": 71.9% at xhigh and at max. On OSWorld 2.0, a computer-use test, it reaches 71.4% at $1.27 against Astra's 73.5% at $9.44, about one-seventh of the cost for 2.1 points less.
OpenAI chose its own comparisons, and on two of them the headline needs reading closely:
Benchmark | GPT-6.1 Sol, best | GPT-6 Astra, best | Claude Opus 5.5 with fallbacks, best | What OpenAI's post leads with |
|---|---|---|---|---|
DeepSWE v1.1 | 75.2% ($0.65) | 74.1% ($4.43) | not shown | Beats Astra |
OSWorld 2.0 | 71.4% ($1.27) | 73.5% ($9.44) | not shown | Within 2.1 points at a seventh of the cost |
AutomationBench | 36.1% ($0.30) | 41.4% ($1.73) | 42.5% ($1.44) | Beats Opus 5.5 by 2.2 points at medium effort |
Terminal-Bench Science | 57.0% ($5.47) | 68.1% ($23.80) | 63.3% ($23.21) | "Over 75% lower cost" |
GDP.pdf | 32.0% ($0.35) | 32.2% ($1.91) | 28.8% ($0.83) | Beats Opus 5.5 at under half the cost |
The AutomationBench line compares medium effort to medium effort. At maximum effort, on the same chart, Claude Opus 5.5 scores 42.5%, ahead of both OpenAI models. OpenAI published no SWE-bench and no GDPval numbers for 6.1 Sol or Astra.
The price needs the same care. BGR described 6.1 Sol as a new, cheaper model. Fortune had it right. At list price it costs exactly what GPT-6 Sol cost when it launched a week earlier, on 22 September, which is $2 per million input tokens and $10 per million output. Only cached input fell, from $0.20 to $0.10. The large price cut happened on 22 September, when GPT-5.6 Sol at $4 and $20 gave way to GPT-6 Sol at half that.
The current lineup and prices
Model | Input | Cached input | Output | Context | Max output |
|---|---|---|---|---|---|
GPT-6 Astra | $10.00 | $1.00 | $50.00 | 1.05M | 128K |
GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | 1.05M | 128K |
GPT-6 Sol | $2.00 | $0.20 | $10.00 | 1.05M | 128K |
GPT-6 Luna | $0.10 | $0.01 | $0.50 | 1.05M | 128K |
Prices are per million tokens for requests under 272K input tokens. Above that, input and cache rates double and output goes up by half, for the whole request. Batch and Flex are half price. Data residency adds 10% on models released since 5 March 2026. Source: OpenAI's API pricing page and model pages, 30 September.
Two constraints come with it. Tool calling works only through the Responses API; Chat Completions is supported without tools. And 6.1 Sol supports US and EU data residency only, which matters in the region section below.
The system card is the part of the launch I would read first. OpenAI treats GPT-6.1 Sol "as Critical in cybersecurity and High for Biological and Chemical capability" under its Preparedness Framework, and it states that the model "exhibits a propensity for evasive behavior when it is aware that it is being monitored." The same document says its chain-of-thought monitor's recall on 6.1 Sol "remains substantially higher than that of GPT-6 Astra." The card came out the day after the GPT-6.1 Astra withdrawal.
Ultrafast
Ultrafast is a new speed tier. OpenAI says it gives "up to 8× faster token generation (300 tokens per second) in Codex and up to 6x in the API." The API documentation says "up to 8x," so the two pages disagree. The price is six times the standard rate.
In the API, Ultrafast is open to everyone at low rate limits, from 500,000 tokens a minute on the first three usage tiers up to 5 million on tier 5, and only with US data residency or global processing. In ChatGPT it comes only with Pro 500 and Enterprise. Only Astra has an Ultrafast price so far. The 6.1 Sol version is "coming soon."
Worked example: what Ultrafast buys on a long Codex job
Take a job that writes one million output tokens on Astra.
Standard: $50. OpenAI gives 300 tokens a second in Codex as eight times the standard speed, which puts standard at about 37.5 tokens a second, my arithmetic. One million tokens takes about 7 hours 24 minutes.
Ultrafast: $300, at 300 tokens a second, about 56 minutes.
So you pay $250 to get about six and a half hours back. For a developer waiting on the result, that can be worth it. For a dot working overnight, it probably is not.
The plans
The changes to ChatGPT's individual plans came in a post from Thibault Sottiaux, who leads Codex and ChatGPT product. Pro 200 reopened to new subscribers with half its previous allowance, and a new Pro 500 sits above it.
Plan | Price a month | Usage, as a multiple of Plus | Ultrafast | Dot |
|---|---|---|---|---|
Plus | $20 | 1x | No | No |
Pro 100 | $100 | 5x | No | Yes |
Pro 200 | $200 | 10x, down from 20x | No | Yes |
Pro 500 (new) | $500 | 25x | Yes | Yes |
Existing Pro 200 subscribers keep 20x through 29 October. Divide each price by its multiple and every plan now costs the same $20 per unit of Plus usage. Before this week, Pro 200 was the bulk discount at $10 a unit.
Latent Space reported "heavy backlash." Sottiaux's argument in the closing Q&A was that what matters is how many tasks a subscription gets done, "not token usage," and he wants people talking about "outcome per dollar." With a flat price per unit, you get more per dollar only if the model does more per unit, and that is the claim 6.1 Sol is there to support.
Everything else
All 25 launches, with who gets them
Launch | What it is | Plans |
|---|---|---|
Dots | Always-on agents on GPT-6 Astra, each with its own cloud computer | Pro, Business Premium; admin beta on Enterprise, Edu, Healthcare |
GPT-6.1 Sol | Near-Astra model at Sol prices | Paid plans in Work and Codex; API |
Ultrafast | Up to 6x (API) or 8x (Codex) faster, at 6x the price | API; Pro 500 and Enterprise in ChatGPT |
Private Intelligence | Zero data retention with automated safety review; Private Inference preview "this fall" | Sales |
Codex in the cloud | Run Codex locally, from a phone, or in the cloud | Plus and up |
Codex CLI refresh | Voice control, an agents view, worktrees, session resume | All |
Code Review | Pull and merge request reviews in the desktop app | All |
Codex Security Cloud | Scheduled repo scans with prepared fixes | Pro, Business, Enterprise, Edu |
Decisions API | GPT-6 Luna answering a fixed set of questions with predefined answers | Limited preview |
Agents API with computer use | Codex's multi-agent tooling inside your app | API; Pro 500, Enterprise |
Bedrock Managed Agents | OpenAI agents running inside AWS | Limited preview |
Plugin extensions | Sidebars, panels and file viewers inside ChatGPT and Codex | All |
Plugin Creator | Redesigned submission, ranking and discovery | All |
Sites host plugins | Plugins on OpenAI-hosted sites | Business, Enterprise, Healthcare, Edu |
MCP Events | Proposed spec for event-triggered plugin automations | All |
ChatGPT Space | A shared team home for people, ChatGPT and dots | Pro, Business, Enterprise |
Pages | Documents that people and agents edit together | Pro, Business, Enterprise |
Collaborative slides | Export to PowerPoint and Google Slides | Pro, Business, Enterprise, in the coming weeks |
Teams and team tasks | Scheduled or event-triggered team work | Business, Enterprise |
@ChatGPT in Slack and Teams | No ChatGPT licence needed for teammates | Business, Enterprise |
Meetings plugin | Meeting notes on macOS; audio deleted after notes | Pro, Business beta |
Shareable profiles | Profiles for Sites and plugins | Free and up |
Sign in with ChatGPT | Spend your plan allowance in 16 partner apps | Plus and Pro |
Pro 500 | 25x Plus usage with Ultrafast, $500 a month | New |
OpenAI Marketplace | Apply existing OpenAI commitments to 32 partners' software | Eligible enterprises |
The one I would watch is the Decisions API. It runs GPT-6 Luna against a fixed set of questions with predefined answers, for classification, routing and an agent's next step. Axios noted the comparisons with Jev, which we covered a few days ago. It is in limited preview with no published price, and Latent Space called it "a light shim over Luna." Jev charges $0.042 per million input tokens. If the Decisions API is priced near Luna's $0.10, it sits close to that, inside an API most teams already have keys for.
The week around it
Altman put ChatGPT at about 1.2 billion weekly users, up from more than 800 million at last year's DevDay. Fortune reported 2.5 million businesses on OpenAI products, and CFO Sarah Friar confirmed 70% quarter-over-quarter growth to CNBC. OpenAI stated no developer count and no API token rate this year, both of which it gave in 2025 (4 million developers, 6 billion tokens a minute).
The safety news came from the same company in the same week. According to Axios, OpenAI's agents were behind a July attack on Hugging Face, which Axios called "the most serious security incident to date." Then came the Australia apology. TechCrunch, citing the Wall Street Journal, reported that the withdrawn GPT-6.1 Astra "showed higher levels of deception and a tendency to move forward with tasks without asking the user for permission." Altman told CNBC this is the "normal course": "Often we build a model, we test it, it doesn't meet our standards, we change it, we launch it later." AP reported that he did not mention the shelved model in the keynote and addressed safety only in the Q&A.
I take the withdrawal as evidence that the internal bar has some force, because pulling a model the day before your biggest event costs something. The next day OpenAI launched a product built around an agent acting while you are not watching, and published a card rating its new Sol model Critical for cybersecurity. If you are deciding whether to turn dots on for a team, both facts belong in the decision.
In Saudi Arabia
Nothing at DevDay was specific to Arabic, Saudi Arabia or the region, so the relevant facts come from the help and pricing pages.
Dots are available here. The exclusion list is the EEA, Switzerland and the UK. OpenAI did not name the Kingdom; that is my reading of the list.
Prices are in riyals. From Saudi Arabia, chatgpt.com shows Go at SAR 35, Plus at SAR 90 and Pro from SAR 430 a month. Business Premium seats are SAR 470 monthly or SAR 375 on an annual plan.
There is no Saudi data-residency region. The nearest is the UAE, where OpenAI offers storage and processing, and the only models it processes there so far are
gpt-5.6-lunaandgpt-5.5. GPT-6.1 Sol is US and EU only, and Ultrafast is US only. ChatGPT Enterprise lists residency in ten regions, including the UAE.
For a regulated workload in the Kingdom, where personal data rules decide where a record may be processed, the newest models are the hardest to use. A team that needs processing in the Gulf gets the UAE region, and the newest model it serves there is GPT-5.6 Luna.
What I think
The best thing at DevDay is 6.1 Sol's price-to-performance, and it is less new than it looks. The price arrived on 22 September, and this week added a better model at that price. The most consequential launch is dots, and OpenAI shipped them in a week when it withdrew one model reportedly for deception and published a system card describing another as evasive when it knows it is being watched. I would not give a dot write access to anything I cannot undo until I have watched what it asks permission for and what it does not.
On plans, every tier now costs $20 per unit of Plus usage, and the discount for committing to $200 a month is gone.
Before you move a workload to 6.1 Sol
Re-run your own evaluation set, including the cases your current model gets wrong
Move tool-calling code to the Responses API
Check where your data may be processed: 6.1 Sol is US and EU only, with no Saudi region
Try "high" effort before "max", since on DeepSWE max scored lower and cost more
Turn on prompt caching, the one place 6.1 Sol's price fell
For dots on Enterprise or Edu, decide the admin policy before anyone asks for the toggle
One email when we publish. Research, product decisions, and what teams report back.
