Sundar Pichai, chief executive of Google and Alphabet, said in his remarks for Alphabet’s third-quarter earnings call, published on Google’s blog on 28 October 2026: “We’re at the point where AI is helping us increase our training and launch velocity. I would expect the next 12 months to see continued acceleration because of RSI.” The remarks do not spell out RSI; in AI research the initials are widely read as recursive self-improvement, AI systems helping to build their successors.
The launch record can be counted. By our count from Google’s release notes and posts, Google released or announced 16 Gemini-branded models between 21 July and 28 October 2026. Thirteen came in the 71 days to 29 September, and 3 in the 29 days from 30 September, when it announced Gemini 4 Argon. Two of those three were Gemini 4 frontier models, Argon and Carbon.
Business Pill · Data Flywheel
An evergreen Business Engineer explainer on data flywheels, when using a product makes it better. It explains the general idea only and does not discuss Google’s remarks.
The key insight: Pichai says AI is now raising Google’s training and launch velocity and expects more over the next 12 months because of RSI, but Google gives no measure of that velocity. By our count, 13 Gemini models launched in the 71 days before Argon and 3 in the 29 days from its announcement, two of them Gemini 4 frontier models.
Pichai said that “at Google, just in the last few weeks, teams across Chrome, Ads, and Search have used Argon to write and optimize code much faster, shipping 35% more code with 30% fewer errors.”
On outside demand, he said that on Google’s developer platforms, Argon’s usage “in the first week ramped up three-and-a-half times faster than Gemini 3, even with capacity constraints.” He said Google has “more models coming soon, including more efficient versions to help developers optimize for speed and cost and to deploy across our billions of users.”
Gemini-branded models released or announced by Google, 21 July to 28 October 2026, by date: 13 in the 71 days to 29 September and 3 in the 29 days from 30 September, when Gemini 4 Argon was announced. Counts are our arithmetic from Google’s release notes and posts; variants counted separately.
What He Added on the Call
Answering JPMorgan’s Doug Anmuth, according to a speech-to-text transcript of the webcast, Pichai said Google first saw a step change internally with Gemini 3.8 Flash and a larger shift once it began deploying Argon internally. He called it one of the most meaningful internal “productivity levers we have seen on software engineering, directly tracking high-quality code that’s getting checked in, the rate at which it’s getting checked in”, and fewer rollbacks.
He said that happened “only about four weeks ago”. He added: “We are now deploying Carbon internally”, and described Carbon as a meaningful step up from Argon. He spoke of a “golden year ahead” in Google’s productivity once it is widely available to Google’s staff with more compute.
What Anthropic’s Filing Says About the Same Loop
Anthropic’s Form S-1, filed with the SEC on 26 October 2026, describes a similar dynamic in its own words: “each new version of a Claude model directly accelerates development of the next.” The filing calls this “closing the loop” and says: “Over 90% of the contributions to our codebase in September 2026 were written by Claude.”
The S-1 also says the average daily volume of code contributed per engineer in the three months ended 30 September 2026 was “over 45 times” the daily average in 2024. Google’s 35% figure and Anthropic’s 90% figure measure different things and are not comparable.
The Launch Record, Counted
By our arithmetic, the 13 launches before 30 September work out to about 5.5 per 30 days, and the 3 launches since to about 3.1 per 30 days. The mix changed: none of the first 13 was a Gemini 4 frontier model, while two of the last three were. Pichai’s expectation is for the next 12 months, and four weeks of releases do not test it either way.
Google announced Argon on 30 September, when it was rolling out to trusted cyber defenders, and said on 13 October that it was “now globally available” to developers and enterprises. Carbon was announced on 28 October; Pichai said it is rolling out “first to a subset of trusted cyber defenders”.
As we read the remarks and the call, Google is describing AI-assisted engineering inside its own model work; the only figure it gives is for code shipped by teams across Chrome, Ads and Search over a few weeks.
Anthropic’s S-1 describes the same kind of loop in its own words and gives a codebase figure; the two companies measure different things, so their numbers cannot be lined up.
Because Google publishes no measure of training or launch velocity, a public launch count is the nearest observable record, and it measures releases rather than how fast models are trained.
“Weโre at the point where AI is helping us increase our training and launch velocity.”
Three Implications
FOR GOOGLE’S ENGINEERING Pichai said teams across Chrome, Ads and Search shipped 35% more code with 30% fewer errors using Argon in the last few weeks.
FOR THE LAUNCH RECORD By our arithmetic, launches ran at about 5.5 per 30 days before 30 September and about 3.1 per 30 days since, with both Gemini 4 frontier models in the later window.
FOR THE MEASUREGoogle gave no measure of its training or launch velocity, so the acceleration Pichai describes cannot be sized from these documents.
The framework starts from two curves: “In any product cycle, there are two curves to follow. The first is what your underlying technology can do. The second is what your users can actually do with it.”
Pichai’s remarks describe the first curve speeding up inside Google, and say Google is bringing agentic features to Search and the Gemini app as a next step. The framework was published in June 2026 and does not discuss Google’s remarks; whether the gap it names is widening at Google is not established by them.
What Is Not Established
Google gives no measure of its training or launch velocity, so the acceleration Pichai describes cannot be checked against a number. The remarks do not define RSI or say what share of Google’s model development AI now performs.
The 35% and 30% figures cover teams across Chrome, Ads and Search over a few weeks; Google gives no baseline, team count or method. A count of model launches measures releases, not how fast models are trained, and it depends on how variants are counted.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Pichai says AI is now raising Google’s training and launch velocity and expects more because of RSI, and Anthropic’s S-1 describes the same kind of loop. Google gave a code figure for some teams but no measure of the velocity itself. By our count, both Gemini 4 frontier models arrived in the 29 days from Argon’s announcement on 30 September, while the overall number of launches in that window was lower than in the 71 days before it.
A note on sourcing. Quotations from Sundar Pichai come from his prepared earnings-call remarks published on Google’s blog on 28 October 2026 and from a speech-to-text transcript of the Alphabet Investor Relations webcast of the call. Anthropic’s lines come from its Form S-1. Launch counts are our arithmetic from Google’s release notes and posts; linked earlier pieces carry their own sources.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
