The model behind the answer now changes every 16 days

On 22 September Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna, 18 days after its GPT-6 Astra. On 30 September Google announced Gemini 4 Argon, which it calls its new frontier model, to a set of trusted cyber defenders through its Fairwind Program; developer access followed on 13 October. On 12 October Meta released Muse Wonder and Muse Wonder Full, which it calls its most capable models. In 2023 the whole industry averaged more than five weeks between flagships.

We logged every notable model release from the major labs since ChatGPT launched on 30 November 2022, and marked the ones that were a lab's new top model when they came out. We call those flagships. There are 59 of them.

This briefing was first published on 23 September 2026 with data to that date. On 15 October 2026 we added the seven releases from the three weeks since, each read from the lab's own announcement: Claude Sonnet 5.5 (28 September), Gemini 4 Argon (30 September), the Mistral Large 4 preview (6 October), Claude Haiku 5.5 and the October update of GPT-6 Sol and Luna in ChatGPT (both 7 October), Muse Wonder (12 October) and Grok Voice 3.0 (14 October). Two of the seven are flagships. The mean gap for 2026 moved from 17 days to 16.

Some of the speed-up comes from labs shipping in smaller steps: Claude Opus went from 4.6 to 4.8 in under four months (5 February to 28 May). The charts count releases but do not measure how much each one changed.

In October 2025 OpenAI said publicly it had set internal goals of an automated AI research intern by September 2026 and an automated AI researcher by March 2028 (TechCrunch). In September 2026 it said it had met the first goal, on its own assessment (Engadget).

If AI is doing more of the research behind each model, the gap between releases has room to shrink further. Our working view is a new flagship somewhere in the industry every week or two during 2027, with version numbers mattering less as updates become continuous. We projected that from the curve above. We have not measured it.

We asked one company's 65 buying questions on 4 June and again on 12 August 2026, and changed nothing ourselves in between. Six answers came back different, roughly 9%. Over those ten weeks all three engines we measure moved a generation, competitors kept publishing and search indexes updated, so we cannot say how much of the movement the models caused (Study 02). The answers moved while the models under them changed.

So a score needs its model and date attached. "We appear in 40% of answers" tells a board very little until it says which model answered, and when. Compare two quarters across a model change and part of what you see is the model.

Tactics aimed at one model carry more risk when models change this often. Muck Rack's May 2026 study found 84% of AI citations come from earned media (Muck Rack, "What Is AI Reading?"), so reviews, publishers and comparison sites make up much of what the engines cite. Work on that material has a better chance of surviving the next release than wording tuned to one model's habits.

As AI systems do more of the research and shortlisting for the customer, more of the decision may happen before the click, and your own site becomes one source among many.

What counts. Notable releases from eight lab groups: OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, a group of other Chinese labs (Alibaba, Moonshot, Z.ai, MiniMax, Xiaomi) and Mistral. We left out minor snapshots, fine-tunes and product features. This is not every model ever released. The long tail of smaller labs and open models is much larger.

What a flagship is. Our judgement of a lab's new top model, or a new model family, on the date it became broadly available. Restricted previews, such as Claude Mythos Preview in April 2026, are listed but not counted as flagships. Where a model reached a small group first and everyone later, we use the later date.

Dates. Every 2026 flagship date was checked against launch coverage or the lab's own announcement. Other 2026 releases come from public release trackers (Evertune, benchr, LLM Gateway), which sometimes disagree by a day or two. Trackers also log more small releases than they did in 2023, so the count of all releases probably overstates the rise. The headline numbers use flagships only, so they do not depend on that count.

Averages. "Every 37 days" and "every 16 days" are the mean gap between consecutive flagships across all labs, for flagships released in 2023 and in 2026 (to 15 October). The lab-level figures are medians.