The model half-life thesis sounds rigorous until you actually count the releases. Paul Kinlan at AI Focus (aifoc.us) did that work on May 18, assembling a dataset of every headline model drop from late 2022 through today across a dozen labs, splitting each vendor into its actual shipping sub-series, and plotting the gaps. His conclusion is plain: release cadence has increased, but no halving curve appears in the data.

That conclusion matters more than the catchy framing it displaces.

Why the metaphor spreads anyway

The half-life construct has a seductive structure. Radioactive decay actually does follow an exponential curve, and Moore’s Law, before it slowed, genuinely described a doubling in transistor density roughly every two years. Both are real phenomena with measured constants. “Model half-life” borrows that scientific credibility without offering a measured constant. When someone says releases are happening “twice as fast every six months,” they are pattern-matching to a rhetorical form, not reporting a finding.

The confusion is compounded by how releases are counted. A frontier lab ships a flagship model, a lite variant, an updated snapshot, and a safety patch in the same quarter. Treating those as four releases in a series that previously had one per year creates a false acceleration signal. Kinlan’s dataset controls for this by splitting vendor output into actual sub-series. Claude Opus and Claude Sonnet are different lines with different cadences. GPT-4 and the o-series have separate release histories. Once you separate the series, the halving story does not hold.

What the data does show

Activity has increased. More labs are shipping more series. Chinese labs, which were largely absent from the competitive picture in 2022, now contribute a meaningful share of drops. The total number of headline releases per year across the full field is higher. But “more activity overall” and “each series ships twice as fast every six months” are not the same claim. The first is true. The second is not supported.

Kinlan’s prediction method, which uses the trailing three gap-medians per series and flags overdue drops explicitly, illustrates the practical implication: individual series still show irregular cadences. A single outlier drop, or a long unplanned hiatus, shifts the median substantially. If a halving law were operating, outliers would average out and the curve would be smooth. The data does not produce a smooth curve.

The operator consequence

The half-life framing is not just intellectually sloppy. It has real downstream effects on how organizations plan around AI infrastructure. Teams negotiating multi-year compute contracts, setting eval refresh cycles, or planning deprecation windows are making financial decisions based on assumed release cadences. If those assumptions are built on “half-life” intuition rather than measured series data, they carry a compounding error.

A company that assumes it must re-evaluate its model vendor every three months because “half-life is shrinking” will over-invest in evaluation infrastructure. A company that anchors on a specific series and tracks its actual gap history will make a more calibrated bet. Kinlan’s dataset, which he plans to update every few months, is exactly the kind of empirical grounding those decisions should rest on.

The comparison that holds

Moore’s Law is instructive precisely because it eventually failed. For roughly forty years, the doubling curve was real and measurable, and the industry organized around it: chip architecture, software design, business model assumptions all priced in the expectation of predictable density gains. When the physical limits of silicon ended the curve, the industry needed new frameworks. The lesson was not that exponential laws govern technology forever. It was that when a measurable law exists, you should use it, and when someone claims one without measurement, you should ask for the numbers.

Nobody has produced the numbers for model half-life. Kinlan looked and did not find them.

The practical stance for operators: track the specific series you depend on, use actual gap history to project the next drop, and discount claims about exponential acceleration unless the person making them shows you a chart with a fitted curve and a confidence interval.

Paul Kinlan published this analysis at AI Focus (aifoc.us) on May 18, 2026.