Plotting AI model release cadence: two labs are accelerating, three aren't

Plotting frontier model release cadence, with methodology

Ethan Mollick made an offhand observation this month: if AI self-improvement is real, even weakly, then the labs that have it should ship faster over time, and the ones that don't should fall behind. He claimed this was already visible at Anthropic and OpenAI but nowhere else. I wanted to check whether the release data actually supports that, so I plotted it.

If AI self-improvement, even in a very limited way, is possible, the cadence of shipping both AI products, harnesses, and models should go up. This appears to be happening at Anthropic and OpenAI, but not for any other labs, including those that seemed to be catching up last year. Ethan Mollick, June 19, 2026 [1]

The claim is falsifiable, which is rare for AI-progress takes, so it's worth testing against data rather than vibes. Here's the cumulative count of major frontier model releases per lab since Q1 2023.

The thing to look at is which lines are bending. Anthropic and OpenAI don't just have the steepest slopes, their slopes increase toward the right. Google sat nearly flat through 2025, then sprinted in Q2 2026. Meta plateaued after Llama 4 in April 2025 and hasn't shipped a frontier model since. DeepSeek runs a steady quarterly cadence without accelerating.

To isolate acceleration, here's the annualized release rate, a trailing four-quarter window. On this view a flat horizontal line means constant cadence; an upward-bending line means accelerating cadence.

Two labs bend up. Three don't. Anthropic roughly tripled its annualized rate over the window; OpenAI more than doubled. Google held flat until a 2026 catch-up; Meta is in decline.

The recursion argument

There's a deflationary reading where this is just spending and headcount, and the cadence gap won't compound. The argument that it does compound rests on three things landing at the same two labs at the same time. None of them proves recursion alone; together they're exactly what you'd expect to see if the loop were real.

1. The labs build their successors with their own products. Anthropic engineers use Claude Code to write the training and eval infrastructure for the next Claude; OpenAI uses Codex to ship Codex. Each release improves the harness that produces the next release, so the next one ships sooner and better — the literal "products, harnesses, and models" Mollick named. Note what this is and isn't: the deployed model is frozen between versions, so there's no online learning inside the weights. The recursion is at the level of the organization, not the forward pass. Call it offline RSI — the loop closes across release cycles, not within them. That's a far weaker claim than "self-improving AI," and it's the one the chart is actually consistent with.

2. Talent is concentrating where the loop already runs. In the week of June 19, 2026, two of the decade's biggest moves landed together: Noam Shazeer — Transformer co-author, architect of multi-query attention and modern MoE routing — joined OpenAI to lead architecture research, and John Jumper — AlphaFold, 2024 Nobel — left Google DeepMind for Anthropic after nine years [5]. The most important living architecture researcher and the most decorated scientist in the field both moved to the two labs already shipping fastest. Talent follows shipping velocity, and shipping velocity benefits from talent. That's a feedback loop, not a coincidence.

3. Compute per cycle is falling fastest where it matters. Tri Dao's FlashAttention-4 hit 71% utilization (1605 TFLOPs/s) on NVIDIA B200 in March 2026 [3], and Mamba-3, from the same group, was the first major architecture release explicitly designed inference-first rather than training-first [4]. Cheaper training and inference per cycle means more cycles per quarter, which feeds straight back into cadence. The efficiency frontier is moving in the same direction as the release frontier.

Any one of these is explainable as money or luck. All three, at the same two labs, in the same window, is the pattern a compounding loop would produce. The case is circumstantial — I'd rather say so than overclaim — but it's coherent, and it's the reading the data is consistent with.

The market read-through (this is the speculative part)

Everything above is defensible from public data. This section is not. It's the trade the chart implies if you believe the loop is real — flagged as speculation up front, so you can stop here if you only came for the data.

The labs themselves are mostly uninvestable: Anthropic and OpenAI are private. So the exposure routes through their compute substrate, their public cloud proxies, and the physical inputs compounding inference consumes.

The cleanest long is NVDA plus the power names, with MSFT as a corollary on the OpenAI loop. The cleanest fade is the cohort that isn't bending: Google equity until its cadence visibly closes the gap rather than catching up for a single quarter, Meta until a frontier model actually ships, and any data-licensing business whose moat is static information rather than environment-coupled learning. One live risk to the MSFT leg: the Cursor "Origin" Git launch is a real threat to GitHub's hold on the dev substrate — the loop runs on GitHub for now, but that isn't guaranteed.

What would falsify this

The honest failure modes, since the whole point was to test a falsifiable claim:

• The slope stops bending. If the next two quarters show Anthropic and OpenAI flattening, the acceleration was a 2026 artifact, not a loop. This is the cleanest test, and it prints soon.
• A lagging lab ships true online learning first. If continual in-weights learning arrives from Google, an open-source effort, or anyone outside the cohort, the offline-RSI cadence advantage stops mattering — the slow labs jump straight to the next paradigm and the moat inverts.
• A paradigm shift lands outside the cohort. If the post-transformer architecture (scaled Mamba-3 systems, pure-RL agents, whatever comes next) ships from Together AI, Cartesia, or an academic lab, the incumbents lose their edge even while shipping at the current cadence.
• Regulation breaks the loop. The week of June 19, Anthropic shipped Fable 5 with terms barring developers from using it to build competing LLMs, and the U.S. Commerce Department issued export controls on Mythos and Fable requiring licenses for foreign nationals; Andrew Ng's open letter argued the stacked restrictions are pushing other nations toward sovereign alternatives [7]. If safety incidents or controls interrupt deployment, the loop stops compounding regardless of the underlying capability.
• The release definition is doing the work. Count point releases and the picture shifts; exclude them, as I did, and you risk understating a lab like Meta that moved to a different release philosophy. The date list is the thing to argue with.

The bottom line

The chart doesn't prove AI is improving, and it doesn't prove recursion. What it shows is narrower and harder to wave away: two specific labs have a release cadence that is accelerating, three don't, exactly as Mollick described — and the same two labs are simultaneously winning the talent and compute-efficiency races the recursion reading predicts. If that's a loop, the gap keeps widening, and the trade is the spread between what's visible in this chart and what's priced once it's consensus. If it's funding or luck, it regresses. Either way the next two quarters of release dates are a clean, dated test, and the prediction is on the record.