It's been a quieter week by 2026's standards — no trillion-parameter launch, no UN session, no CEO signing a public letter about human extinction. But underneath the quiet, two of the industry's biggest labs made moves that say more about where this is all headed than another benchmark chart would. OpenAI told developers it now has 1.2 billion people using ChatGPT every week. Anthropic, instead of answering with a bigger model, opened something closer to a trade school.
Add in a strange, barely-noticed Google Research paper that used Gemini to prove genuinely new mathematics, and the week from September 29 to October 3 turns out to be less about new chatbots and more about what happens once a technology this size has to actually scale into institutions — enterprises, schools, and now, apparently, open research problems.
Quick summary: OpenAI's DevDay 2026 (Sept 29) revealed ChatGPT now reaches 1.2 billion weekly users and shipped GPT-6.1 Sol, a coding-and-agents model priced at one-fifth of flagship GPT-6 Astra rates. On October 2, Anthropic launched Claude Frontier Academy, a $100 million, nomination-only program aiming to train 10,000 "Frontier Deployed Engineers" by the end of 2027, starting with 100 engineers from Accenture, Deloitte, McKinsey, Morgan Stanley and others. Separately, a Google Research preprint posted September 30–October 1 describes Cogentic, a multi-agent system built on Gemini that produced expert-verified new results on five open problems in online learning, auction theory and mechanism design — covered in more depth in a separate post.
OpenAI's DevDay: a cheaper near-flagship, and the user number everyone was waiting for
OpenAI's annual DevDay landed September 29 at Fort Mason Center in San Francisco, and the headline product was GPT-6.1 Sol — an upgrade to the mid-tier GPT-6 Sol model aimed squarely at agentic coding, computer-use tasks, and professional work. The pitch is specific: near-GPT-6-Astra intelligence at roughly one-fifth of Astra's standard per-token price. Confirmed pricing is $2.00 per million input tokens and $10.00 per million output tokens, with cached input dropping to $0.10 per million tokens, a 1.05-million-token context window, and up to 128,000 output tokens. It's live now via the API, ChatGPT Work, and Codex (Sources: InfoQ, byteiota).
Alongside it, OpenAI expanded its Agents API with native computer-use and multi-agent support, launched "Dots" — persistent, always-on agents for ongoing task management — rolled out Codex Cloud for running coding tasks remotely across devices, and previewed a Decisions API that routes requests to predefined answers using the cheap GPT-6 Luna model under the hood. OpenAI is also taking DevDay international, with eight "DevDay Exchange" events planned across Asia, Europe and Latin America through October and November. The number that got the most attention, though, was usage: OpenAI told developers ChatGPT now reaches 1.2 billion weekly active users, up from roughly 900 million in February and just over 1 billion as of late August — a genuinely fast climb for a two-month stretch (Sources: OpenAI DevDay recap, TechFlow Post).
Anthropic didn't ship a model. It opened a school.
Three days later, on October 2, Anthropic announced the Claude Frontier Academy — a $100 million program built to train 10,000 "Frontier Deployed Engineers" by the end of 2027. It's not open enrollment: participants have to be nominated by their employer through Anthropic's enterprise account team or a partner network, and the company says it's looking for solid software engineers with a track record on language models rather than AI specialists specifically. The structure borrows explicitly from medical residency: a multi-day, in-person phase simulating a full enterprise Claude deployment — including the security review and handover — followed by a 12-week residency in which the engineer leads a real Claude use case inside their own company, backed by Anthropic engineers and peer review. The first cohort of 100 is drawn from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk (Sources: Yahoo Finance, Cybernews).
Read next to OpenAI's user-count flex, the contrast is useful: one lab is optimizing for reaching more individual people more cheaply, the other is betting that the actual bottleneck to AI adoption at the enterprise level isn't model quality anymore — it's the shortage of people who know how to actually wire Claude into a bank's or consultancy's existing systems without breaking something.
| Story | Company | Date | What it actually means |
|---|---|---|---|
| GPT-6.1 Sol launch | OpenAI | Sept 29 | Near-flagship coding/agent model at ~20% of Astra's price |
| 1.2B weekly users disclosed | OpenAI | Sept 29 | Up from ~1B in late August — fastest growth window yet |
| Claude Frontier Academy | Anthropic | Oct 2 | $100M bet that enterprise deployment, not model quality, is the bottleneck |
| Cogentic preprint posted | Google Research | Sept 30–Oct 1 | Gemini-based system proves new results on 5 open research problems |
A quieter, stranger thread: Google's AI started proving new math
Easy to miss amid the product launches: a Google Research preprint posted right at the turn of the month describes Cogentic, a multi-agent system that used Gemini to produce new, expert-verified results on five genuinely unsolved problems in online learning, auction theory, and mechanism design — not textbook exercises, but open research questions nobody had answered before. It's different enough from the usual "AI passes a math benchmark" story that it gets its own full breakdown elsewhere on LumiChats, but the short version is: this isn't AI checking its own homework anymore, it's AI work that independent mathematicians had to actually sit down and verify.
If you're evaluating OpenAI vs. Anthropic for a business right now, this week is a useful signal of each company's actual bet: OpenAI is optimizing for reach and price at the mid-tier (GPT-6.1 Sol), while Anthropic is optimizing for depth of enterprise deployment (Frontier Academy). Neither is obviously "ahead" — they're solving different problems.
What this week actually adds up to
None of these three stories individually is a blockbuster. Together, they describe an industry that's stopped treating "ship a bigger model" as the only lever worth pulling. OpenAI is chasing distribution and price. Anthropic is chasing institutional integration deep enough that it needs a residency program, not a sales deck. And Google just quietly demonstrated that the same infrastructure can be pointed at problems nobody — human or machine — had solved before. That's a wider spread of strategies than the headlines from earlier in September suggested, and it's worth watching which bet actually pays off over the next few months.
01How many people use ChatGPT every week as of October 2026?
OpenAI said at its September 29 DevDay that ChatGPT reaches 1.2 billion weekly active users, up from about 1 billion in late August 2026 and roughly 900 million in February 2026.
02What is GPT-6.1 Sol and how is it priced?
It's an upgraded version of OpenAI's mid-tier GPT-6 Sol model, aimed at coding, computer-use, and professional agent tasks. Confirmed pricing is $2.00 per million input tokens and $10.00 per million output tokens (cached input $0.10/million), with a 1.05-million-token context window — about one-fifth of flagship GPT-6 Astra's standard rate for comparable performance on several evaluations.
03What is Anthropic's Claude Frontier Academy?
A $100 million, nomination-only training program announced October 2, 2026, aiming to certify 10,000 "Frontier Deployed Engineers" by the end of 2027 through an in-person training phase plus a 12-week enterprise residency. The first cohort of 100 engineers comes from firms including Accenture, Deloitte, McKinsey and Morgan Stanley.
04Is Google's Cogentic the same as DeepMind's AlphaProof?
No — they're different systems. AlphaProof Nexus and Aletheia (also Gemini-based) have been used to resolve a number of unsolved Erdős problems and OEIS conjectures. Cogentic is a separate multi-agent orchestration framework that produced new verified results specifically in online learning, auction theory and mechanism design, described in a preprint posted at the end of September 2026.
Whichever side of this week's split you find more interesting — OpenAI's reach-and-price strategy or Anthropic's enterprise-depth bet — the fastest way to form your own opinion is to actually use both. LumiChats gives you GPT-6.1 Sol, Claude, Gemini and 40+ other leading models side by side in one subscription, so you can judge for yourself instead of just reading the press releases.
