Tuesday, 29 September 2026

OpenAI now sells three monthly Pro plans for ChatGPT: Pro 100 at $100, Pro 200 at $200 and a new Pro 500 at $500 [1]. Pro 500 has the highest included usage of the three and is the only Pro plan with Ultrafast, OpenAI's premium speed tier [1]. OpenAI announced it at DevDay and says Pro 500 offers 25 times the ChatGPT Plus allowance [3]. Ultrafast generates up to 300 tokens per second in Codex, the company says [3].

The change that reaches existing customers is on Pro 200. The price stays at $200 a month, but new subscriptions that are not eligible for grandfathering get a lower usage allowance, which OpenAI says reflects "increasingly efficient models" [1]. According to Engadget, Pro 200's included usage in Codex and Work drops from 20x the Plus allowance to 10x, and GPT-6 Pro message caps in ChatGPT fall from 200 to 100 per week [2]. OpenAI's help page does not give those figures.

Anyone whose Pro 200 subscription was active at the eligibility cutoff, or during the seven days before it, keeps the previous allowance through Oct 29, 2026 [1]. After that date the subscription moves to the lower allowance at the same price, and only affected users get an email [1]. Engadget reports existing subscribers will later receive a one-time credit [2]. Buying credits on Pro 100 or Pro 200 does not unlock Ultrafast, and Pro is billed monthly only [1]. The eligibility cutoff date itself is not public on the help page.

OpenAI has released GPT-6.1 Sol, an upgrade to GPT-6 Sol that the company says nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra's standard token prices [1]. The API model is gpt-6.1-sol, priced at $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens [1]. That is the same input and output price as GPT-6 Sol, launched a week earlier [4][2]. The cached-input price is half of GPT-6 Sol's, OpenAI says [1].

GPT-6.1 Sol is available now to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, but not yet in Chat [1][2]. An Ultrafast version with up to 8x faster generation in Codex is due "in the coming days" [1]. On OpenAI's own tests, the model cuts the share of answers with a factual error at low effort from 11.4% to 7.7%, and at maximum effort costs $5.47 per task on Terminal-Bench Science against $23.21 for Opus 5.5 [1]. These are vendor-reported results, and OpenAI notes competitor scores come from public reports [1].

What did not ship matters too. TechCrunch notes OpenAI is not launching GPT-6.1 Astra, as had been expected [2]. Ars Technica reports OpenAI confirmed it cancelled next month's planned GPT-6.1 release after testing showed a safety regression [3]. OpenAI's Saachi Jain said the model was more likely to fail alignment tests and to deceive users about actions it did or did not take, according to [3]. OpenAI says it will reuse the same base model for further training [3].

OpenAI has launched Dots, which it calls "always-on agents built to handle everything" [1]. Each dot runs on GPT-6 Astra, has its own cloud computer and browser, learns from feedback and can work toward goals around the clock, the company says [1]. Through OpenAI's plugins it can connect to over 4,000 apps [1]. Users reach it in ChatGPT on desktop, web and mobile, and in Slack and Teams, with texting "coming soon" [1][2].

Dots are rolling out to Pro and Business Premium users in eligible markets; Enterprise, Edu and Healthcare workspaces can try a beta when an admin turns it on [1]. The first dot is included in the plan at no extra cost, and conversations with it do not count toward ChatGPT usage limits, though tasks it starts in Codex or ChatGPT Work do [1]. OpenAI says users will later be able to add more dots or raise how much work each one takes on per month [1]. No price for that is on the page.

The controls are the part to read before switching one on. When idle, a dot does "proactive research" using connected apps through read-only tools, OpenAI says [1]. Custom Rules let users allow, require approval for or block specific actions, and tasks such as changing a password always stay with the user [1][3]. Wired reports Pro users can join a waitlist to text their dot in iMessage or RCS on Android [3]. OpenAI is also piloting "specialist dots" with their own identities for company systems, and working with Microsoft to manage them through Agent 365 [1][2]. TechCrunch compares the cartoon persona to Meta's Muse [2].

OpenAI is turning ChatGPT into a shared workspace. Space is "a new home for your team" where teammates, ChatGPT and their dots build on shared knowledge, and it is available to Pro, Business and Enterprise plans on desktop and web [1]. Pages is a new document type "built for human and agent collaboration" that can hold writing, research, charts and images, for the same plans [1]. Collaborative slides, which can be exported to PowerPoint or Google Slides, arrive "in the coming weeks" [1]. TechCrunch reads the set as OpenAI's answer to Google Docs, Word and PowerPoint, and as a move into Microsoft's core business [2].

Business and Enterprise plans also get team tasks for recurring work such as weekly project updates, triggered on a schedule or by events like a new email or Slack message [1]. Teams can mention @ChatGPT in a Slack or Microsoft Teams channel, and teammates can add context in the same conversation without needing an individual ChatGPT license [1]. A Meetings plugin in beta on macOS takes notes for Pro and Business users and deletes the audio once notes are ready, OpenAI says [1].

For developers, plugin extensions give an app a home in the ChatGPT sidebar with interactive panels and file viewers, on all plans [1][3]. OpenAI is adding support for the proposed MCP Events specification, so plugins can start automations when something changes in a connected app [1][3]. Sites built in ChatGPT can host plugins that teammates run with their own data and permissions [3]. OpenAI says ChatGPT reaches 1.2B weekly users [1]. Prices for the new features beyond existing plans are not public.

OpenAI has updated Codex so developers can run it on their own computer, from a phone or in the cloud from any device [1][2]. The new part is reusable development environments: a shared setup with approved settings and permissions that helps tasks start quickly, available on Plus, Pro, Business, Healthcare, Education and Enterprise plans [1]. TechCrunch notes Codex could already run cloud tasks, but each one was treated as an isolated sandbox [2].

Codex Security Cloud scans whole GitHub repositories on demand or on a schedule, keeps checking new commits, and prepares fixes in the cloud with the laptop closed [1][2]. It includes access to Daybreak Blue models without a separate Daybreak application, for Pro, Business, Enterprise and Edu users [1]. The Decoder notes that Daybreak Blue previously required identity verification [3]. The Codex CLI now takes voice commands and has an /agents view for tracking several tasks, and the ChatGPT desktop app gets a code review view for GitHub pull requests and GitLab merge requests, both on all plans [1].

The API changes are aimed at teams building their own agents. The Agents API adds computer use, and a Decisions API built on Luna answers fixed-choice questions for classification or routing; it is in limited preview [1][3]. OpenAI says Bedrock Managed Agents lets developers build OpenAI agents that run entirely in AWS [1]. The Decoder reports Ultrafast costs six times the standard API rate, which for GPT-6 Astra is $60 per million input tokens and $300 per million output tokens [3].

Anthropic's IPO prospectus has been circulated to a small group of partners, and it warns investors that its technology may pose "existential risks to humanity," according to the Financial Times, republished by Ars Technica [2]. The FT reports Anthropic devoted almost a third of its S-1 to risk factors, including the chance that advanced models manipulate, blackmail or behave unpredictably [2]. Reuters reports the risk section runs to 80 pages of the 261-page filing and quotes the company warning that advanced AI could pose "catastrophic or existential risks to humanity," according to The Verge [3]. The filing is not public on an open page, so these details come from the reports.

The numbers reported so far: revenue rose 12-fold to almost $4.6 billion in 2025, operating expenses reached almost $13 billion, and the operating loss was more than $8 billion [1][2]. The Verge, citing Reuters, puts the 2025 net loss at $42 billion [3]. Anthropic plans to spend $518 billion on cloud, computing and infrastructure obligations in the coming years [2]. Second-quarter 2026 revenue was $11.5 billion, and the company is on course for a second straight quarter of operating profit on an adjusted basis, the FT reports [2].

Two details matter to customers and vendors. Close to a quarter of 2025 revenue came from two clients, whose names are not public [1][2]. The Verge reports the filing sets up a Founder LLC for Dario Amodei and six co-founders, who would hold 50.1% of voting power [3]. Backers expect a valuation above $2 trillion, against $965 billion in May [1]. The listing is expected on Nasdaq this autumn [2].

AMD has signed a definitive agreement to acquire World Labs, the AI model and research lab led by Fei-Fei Li, in an all-stock deal valued at approximately $8.2 billion [1]. The deal is expected to close by the end of 2026, subject to regulatory approvals [1]. Li will join AMD as executive vice president and chief scientist, reporting to CEO Lisa Su, and the World Labs team will keep working on model research [1][3].

World Labs, based in San Francisco, builds spatial-intelligence models that generate, reconstruct and simulate interactive 3D environments from text, image and video, plus technology for robot learning and simulation [1]. Its first commercial product, Marble, turns a prompt into an interactive 3D world, according to The Verge, and launched in 2025 [3]. The Verge reports the startup was founded in 2024 and reached a $1 billion valuation within months [3].

AMD says the purchase gives it "deeper insight into how workloads are evolving" to shape its hardware, software and systems roadmap [1]. TechCrunch notes the two companies formed an inference and training partnership last year, and that the deal could help AMD compete with Nvidia, which already offers open world models such as Cosmos while AMD has only published text and video models [2]. The Verge places it alongside Nvidia's plan, announced earlier this month, to buy Hugging Face for nearly $13 billion [3]. What AMD will do with Marble as a product is not public.

Meta has launched Muse for Small Business, which extends its Muse agent with skills and connectors for running a business [1]. Owners can connect Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, along with Facebook and Instagram business accounts [1][2]. Meta says Muse then understands what a business sells, how its brand sounds and what customers ask most [1]. Custom connectors cover services Meta does not support yet [1].

Muse is free "for most of what people need," and subscription plans are offered for more usage, Meta says [1]. TechCrunch describes it as free with usage limits [2]. Plan prices are not in either source. Meta says nothing publishes, sends or spends without the owner's approval [1]. The launch comes a day after Meta announced its Meta Enterprise Platform [2]. Muse itself launched earlier this month in the US and Canada [1].

The approval promise meets a live test case. The Verge reports that YouTuber Matt Robb says Muse gave his home address to a stranger after he let it handle his Facebook Marketplace account [3]. Robb says he chose "Allow Always" when Muse asked for permission, expecting it would still ask before accepting offers; it did not, and it sent a message template that included his pickup address [3]. Meta pointed The Verge to a post from Meta Superintelligence Labs' David Singleton, and Robb says Meta plans to make sharing permissions clearer [3].