How often each AI provider ships

What counts as a release. Every row on a per-family Versions page with a parseable ship date counts as one release. The version pages are the canonical record of what shipped and when; this page rolls them up. There is no benchmark- or capability-aware weighting — every release counts equally.

Per-provider grouping. One track per provider. OpenAI's GPT line and o-series roll into one OpenAI track; Anthropic's Opus / Sonnet / Haiku tiers roll into one Anthropic track; Google's Gemini Pro / Flash variants roll into one Google track. The per-tier and per-line nuance lives on the per-family pages.

Rolling-12-month window. The headline number, the "Last 12mo" column, and the "Median 12mo" column all use a trailing 365-day window anchored to the most recent ship date in the dataset. Anchoring to the data's own most-recent date (rather than to today) keeps the page deterministic between refreshes.

Median interval. The median number of days between consecutive sorted ship dates. Median (rather than mean) is more robust to the long single-release gaps that punctuate every provider's history.

What is intentionally excluded. Predictions about future cadence. "Is this sustainable?" editorial commentary. Model-quality weighting. Per-tier and per-line breakdowns within a provider. Benchmark-aware "frontier-pushing" filters. Every release counts equally, and the page measures rate of shipping, nothing else.

Last updated: 2026-05-09. Data anchor: 2026-04-24.