How AI Labs Version Their Models

Updated

A model name such as Claude Haiku 5.5 or GPT-6.1 Sol packs three things into a few words: the family, the tier inside it, and the generation. The API id adds a fourth, which exact build you get. Knowing which part changed tells you how big a release is.

Family, tier and generation

Most labs ship a family of models at once and split it into tiers by size, price and speed. Anthropic uses Opus, Sonnet and Haiku, from largest to smallest. Google uses Pro, Flash and Flash-Lite for Gemini. OpenAI has used suffixes such as mini and nano, and lately named variants.

The number is the generation. A whole-number jump (4 to 5) usually means a new base model with a new training run. A point release (5 to 5.5) usually means the same generation improved, with better post-training, a longer context window or lower prices, and the tiers often update on different days.

Dated snapshots and aliases

Behind each marketed name is a fixed snapshot. Labs have long given snapshots dated ids, for example gpt-4o-2024-08-06 or claude-3-5-sonnet-20241022, and a short alias that points at the newest snapshot of that model.

The difference matters in production. Pinning a dated id means your output stays the same until that snapshot is retired. Calling an alias means you get improvements automatically, and also behaviour changes you did not test. When a release on this site says a model was updated without a new name, it usually means the alias moved to a new snapshot.

Preview, experimental and latest

Labs also publish ids that are not meant to stay: preview, experimental or exp builds, and ids ending in latest. These let people try a model before it is final, and they can change or disappear on short notice. Our guide to release stages explains what each label promises.

Same model, many places

One model often launches on the lab's own app and API first, then on cloud platforms such as Amazon Bedrock, Google Cloud Vertex AI or Microsoft Foundry, and inside other products such as Copilot, Cursor or Replit. Each arrival is a separate announcement, so the same model can appear several times in a month across different company pages. We list each one under the company that announced it.

Reading a release on this site

Releases in the Model category are new models, new snapshots, or changes to how an existing model behaves or is priced. To see how big a release is, check which part of the name changed: a new generation, a new tier, or only a dated id. Then check where it is available, because a model that is only in a research preview is not the same as one that is generally available on every platform.

Or browse the release logs for every AI company we track.