The month the version number stopped meaning anything

Five weeks, three version numbers

Google shipped Gemini 3.7 Flash roughly three weeks after 3.6 Flash, then 3.8 Flash on September 2. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1. Sakana AI released Fugu Ultra v2.0 on September 11. Trackers counted eleven new models from seven providers in the first eleven days of the month, on top of a late-August cluster that included Z.ai's GLM-5.3-Flash and Alibaba's Qwen3.8-Flash-Next on the same day.

At that cadence the version number has quietly changed jobs. Nobody reading 3.8 can infer how much bigger it is than 3.7, because the gap between them was measured in weeks rather than in architecture. The number is no longer a summary of the distance travelled. It is mostly a way of saying which one is newest.

Version numbers used to be a promise

For most of software history, the integer meant something specific and expensive. A major version implied a break: new file formats, retraining, a migration project. Semantic versioning later made that folklore explicit, with the three-part number encoding exactly what could break and what could not. The number was a contract with the people downstream.

The most elegant expression of that idea came from Donald Knuth, who decided TeX's version numbers would converge on the digits of pi, and METAFONT's on e. Each release adds one more digit and gets asymptotically closer to a value it will never reach. It was a deliberate statement that the software was finished, and that from here on there would only be corrections. A version number, in other words, can carry an argument about the nature of the work.

Then the release train replaced the release

The shift began with shipping cadence, not with AI. When Chrome moved to a six-week release train in 2010, the version number stopped tracking significance and started tracking the calendar, which is why Chrome's major version is now a number nobody bothers to memorise. Ubuntu had already made the logic literal by numbering releases after the year and month they shipped. Microsoft went further and declared Windows 10 the last version of Windows, on the theory that a continuously updated product does not have versions at all.

In each case the same trade was made. Predictable cadence is worth more to users than legible magnitude, because a fixed train means a feature that misses this release catches the next one in weeks rather than years. The cost is that the number stops telling you anything, and the release notes have to do all the work.

What a fast number can and cannot tell you

The useful questions have moved to where the number never reached anyway. What changed at the API boundary. What got cheaper. What quietly broke. Anthropic's September 1 release is a clean illustration: it shipped at an unchanged list price and carried three breaking API changes. Neither of those facts is recoverable from the digits, and both are what an integrator actually needs to know.

Price is doing some of the signalling that version numbers used to do. Google introduced 3.7 Flash at half the per-million-token cost of 3.6 Flash and held that introductory price through 3.8. When the number cannot express magnitude, the price sheet and the changelog become the real version history, which is roughly where the industry has landed.