A field survey of version numbers · 2018 to 2026

After the Decimal

Fourteen frontier AI labs, 109 numbered flagship releases since 2018. The integers are inevitable. The decimals are choices, and the industry keeps making the same one.

.5
the industry's favorite decimal
22 of the 60 point releases ever shipped
Scope
Chart 1

Every version, one number line

Each dot is a distinct version number shipped in a lab's primary flagship line, placed at its numeric value. Read the columns: dots stack up over the integers and the halves, and the x.9 band has never been touched.

integer release .5 release other decimal rumored

Grok 4.20 is plotted at 4.2 (the version xAI shipped instead of 4.2, as a 4/20 joke). Hover any dot for the release and its story.

Chart 2

What comes after the dot

The same releases, bucketed by their decimal. Integer launches (.0) are the default state of the world; among true point releases, .5 wins by a landslide, .1 is the runner-up, and the digits past .5 only appear when a lab is counting its way toward the next integer.

integer the half-step everything else

The only .4 ever shipped is GPT-5.4. No Chinese lab has ever shipped a .3 or a .4.

Chart 3

Version velocity

Version number against time, one panel per line, all on the same axes (2018 to 2026, versions 0 to 7). Slope is marketing cadence, not capability: GPT took 86 months to count from 1 to 5, GLM did it in 35, and MiniMax runs at two majors a year.

The signal in the digit

Why .5?

DecimalWhat it signalsShippedExamples
.0A new generation, full launch treatment49GPT-5, Gemini 3, GLM 5, Claude 4
.1Polish: same generation, tightened up12GPT-5.1, Claude 4.1, Gemini 3.1, ERNIE 5.1
.2 to .4Bookkeeping: rapid iteration, rarely a headline10GPT-5.2, DeepSeek V3.2, Llama 3.3
.5Half a generation: launch-worthy, headroom kept22GPT-3.5, Claude 3.5, Gemini 2.5, Qwen 2.5
.6 to .8The ratchet: counting up while the integer bakes16Claude 4.6 to 4.8, Qwen 3.6 to 3.8, Gemini 3.6
.9Never used: it would admit the integer is late0none, in eight years

Your intuition is right, and the data is unambiguous: some decimals are simply more appealing, and .5 is the charm price of AI. It is the only decimal that reads as a quantity ("half") rather than a count. A .5 promises a real leap while explicitly reserving the next integer, which makes it the perfect launch number: big enough for a keynote, humble enough to keep the powder dry. It is no accident that the release that started this whole era, the original ChatGPT, ran on a .5 (GPT-3.5), or that Gemini spent its first two years shipping nothing but .0 and .5. Even Google's fall from that discipline proves the point: when it finally shipped a 3.1 in early 2026, the very next number it reached for was 3.5.

The digits past .5 tell a different story. Once a lab ships a .5, it almost never goes back: it counts .6, .7, .8 in public while the next integer trains. Claude walked 4.5, 4.6, 4.7, 4.8 and then jumped to 5; Qwen is at 3.8 on the same ladder right now; Kimi did 2.5, 2.6, 2.7 before K3; Gemini, fresh off its .0/.5 diet, went 3.5 to 3.6 in two months. And then every single lab stops. Nobody, ever, has shipped a .9: a .9 says "the integer exists and is late," which is the one thing a marketing team cannot say. So .8 is where version numbers go to die.

There is a cultural fingerprint in the data too. Chinese labs love the .5 even more than American ones (12 of their 29 point releases, against 10 of 31 in the US), and they have never shipped a .3 or a .4. The only .4 in the entire dataset is American (GPT-5.4). With 4 a homophone of "death" in Mandarin, tetraphobia is the obvious suspect, and GLM skipping straight from 5.2 to a rumored 5.5 fits the pattern, though with counts this small it stays a hypothesis rather than a verdict.

The deeper shift is what the decimal means. Semantic versioning gave software minor versions as bookkeeping; model marketing turned them into a psychological instrument. A modern lab ships six releases a year but cannot credibly claim six generations, so the decimal became a damper: OpenAI put out 5.1 through 5.6 in eight months, each one news, none of them claiming to be GPT-6. The number after the dot is not a measurement. It is a promise about how impressed you should be, and .5, like $9.99, is the number the industry has learned we like hearing best.

Field notes

Ten naming stories the data can't show

o1 → o3

OpenAI skipped o2 to avoid a trademark clash with the UK telecom O2. Altman said the skip was "out of respect," adding that OpenAI has "a tradition of being very bad at names."

3.5 → 3.7

Anthropic never shipped a Claude 3.6. Users had informally named the October 2024 "Claude 3.5 Sonnet (new)" refresh "3.6," so the next real release jumped to 3.7 to dodge the collision.

4.1 after 4.5

GPT-4.1 shipped two months after GPT-4.5. The numbering ran backwards because 4.1 was the cheaper API workhorse and 4.5 the short-lived giant, deprecated from the API within months.

4.20

xAI followed Grok 4.1 with 4.20, a deliberate 4/20 joke in the versioning. The next release was 4.3, which is numerically a downgrade if you read .20 as twenty. Nobody at xAI seems bothered.

Gemini's broken diet

For two years Google ran the strictest decimal diet in AI: 1.0, 1.5, 2.0, 2.5, 3 and nothing else. Then 2026 happened: 3.1 in February, 3.5 at I/O, 3.6 in July, and a teased Gemini 4. A 3.2 Flash even surfaced in A/B tests in May but never shipped; the pull of .5 won.

5.2 → 5.5?

Zhipu's GLM has run a two-month cadence since mid-2025 (4.5, 4.6, 4.7, 5, 5.1, 5.2). The rumored next step skips 5.3 and 5.4 entirely and lands on 5.5.

M2.5 → M2.7

MiniMax skipped M2.6 outright, going M2.5 then M2.7 five weeks later. No explanation was given, which is itself a data point about how loose these numbers are.

R2, unreleased

DeepSeek's R1 was 2025's biggest release, but R2 never shipped: reporting says founder Liang Wenfeng was unsatisfied with it, and the reasoning line quietly folded back into V3.1's hybrid mode.

4o, X1, T1, K2 Thinking

When a lab wants a launch without spending a number, it reaches for letters: GPT-4o, ERNIE X1, Hunyuan T1 and TurboS, Kimi K2 Thinking, plus the entire Turbo, Max, Plus and Pro economy.

Llama 4 → Muse Spark 1

Meta ended the open Llama line at 4 and relaunched closed-weight as Muse Spark in April 2026: the rare version reset, spending the credibility of a "1" to signal a clean break. It reached 1.1 in July.

Appendix

The full ledger

All 135+ entries, including the unnumbered oddballs and never-shipped ghosts the charts exclude
Methodology
  • One entry per distinct version number in each lab's primary flagship line. Tiers under one number (Opus/Sonnet/Haiku, Pro/Flash, Pro/Lite/Mini) count once.
  • Side lines and unnumbered releases (o-series, DeepSeek R, GPT-4o, ERNIE X1, PaLM, abab, Yi Large) appear in the ledger but not in the counts.
  • Dates are month of first public availability, previews included. Some Chinese-lab dates rely on secondary press.
  • Grok 4.20 is counted as a .2. Rumored releases (GLM 5.5, Grok 5) are excluded unless toggled on.
  • Compiled July 19, 2026; re-audited July 21 with one independent verification agent per model line after a reader caught a missing Gemini 3.1. Dates verified against the sources below.
Sources