An AI model name encodes three things: the family (who made it and what tradition it comes from), the generation or version (which rung of capability and date it belongs to), and the tier (fast-and-cheap versus large-and-capable within that generation) — so "Claude Sonnet 4.5" reads as Anthropic's mid-tier line, generation four, point-five revision. Learn the grammar once and every future name decodes itself, which matters because vendors switch default models silently and the name is often the only visible signal of what changed under your subscription.
This is an explainer built from the vendors' own naming and documentation conventions as of late 2025; version numbers move quickly, so treat specific examples as illustrations of a system, not a catalog. RechargeMe publishes information, not purchase advice.
What does the family name tell you?
The first word is the vendor's product family: GPT for OpenAI's generative pre-trained transformer line, Claude for Anthropic's models, Gemini for Google's, Llama for Meta's open-weight family, and so on. Family mostly tells you about ecosystem — which apps, APIs, and data-use terms come along — rather than ranking. A model's family predicts where you will encounter it: GPT and Gemini are embedded in mass-market consumer apps, Claude leads in certain developer and enterprise channels, and open-weight families like Llama mostly reach you indirectly, running on other companies' infrastructure.
What does the version number mean?
Roughly: the generation, then the revision. A whole-number bump (3 → 4 → 5) marks a new generation — a separately trained system with a step change in capability, per each vendor's release documentation. A decimal bump (4 → 4.5) marks a substantial mid-generation revision. Suffixes like "Turbo" or date-stamps in API identifiers (the year-suffixed model strings vendors document for developers) mark smaller refreshes — the same generation, retrained or retuned, sometimes with changed context limits or pricing. The suffix names are marketing vocabulary, not standards; OpenAI's own documentation has historically used "Turbo" and mini-class names for its cheaper, faster variants, while Anthropic uses Opus, Sonnet, and Haiku as tier names.
What do the tier names mean?
Tier is the most decision-relevant layer. Vendors publish tiers that trade capability for speed and cost: Anthropic's Opus (largest, most capable), Sonnet (balanced), and Haiku (fast, inexpensive) are the clearest-documented example; OpenAI's mini and Turbo variants play the same role on its side; Google's Flash and Pro labels do it for Gemini. Per the vendors' pricing pages, the gap between tiers within one generation can be an order of magnitude in price — which is why routing everyday questions to the fast tier and reserving the top tier for genuinely hard work is one of the few uncontroversial savings in AI tooling. Many consumer and developer products now do this routing automatically, per their release notes, picking a tier per query.
| Layer | Example | What it tells you |
|---|---|---|
| Family | GPT, Claude, Gemini, Llama | Vendor, ecosystem, where you'll meet it |
| Generation | 4 → 4.5 → 5 | Capability rung and training vintage |
| Tier | Opus/Sonnet/Haiku; Pro/Flash; mini/Turbo | Speed and price vs. capability within a generation |
Related stories: How to read a model card: the five sections that carry real information · What AI benchmarks actually measure — and the three ways they mislead.
Why do the names keep changing under you?
Because vendors deprecate aggressively. Model documentation pages routinely mark older versions for retirement, with developers migrated to successors by date — a model that was the flagship eighteen months ago can become unavailable via API entirely. For a subscriber, the practical consequence is that "the chatbot got better" or "the chatbot got weirder" often coincides with a silent default-model switch visible only in a settings screen or release note. The habit worth building: check which model name you are actually talking to, and note it when output quality matters, because it is the strongest single predictor of why results changed between two attempts.
Do bigger numbers mean better for your task?
Usually stronger on average, but not uniformly. Vendors' own model cards document that newer and larger models improve aggregate benchmark scores while still carrying the same core limitations — hallucination among them — and independent evaluation initiatives have repeatedly found that ranking can shift by task type. A top-tier model is a stronger writer and reasoner on average, and it can still fabricate a citation with full confidence. Version literacy tells you what you are using; it does not license trusting it unconditionally.
What the naming systems don't tell you
Parameters, training data, or true cost. The numbers in consumer names stopped corresponding to parameter counts years ago; none of the major vendors discloses training-set composition in its marketing names, and "5" versus "4.5" conveys no unit of measurement. The name is an address, not a specification — the model card and pricing page are where the actual facts live. For independent evaluation efforts that cut across vendor marketing, the journalism at publications like the BBC's technology desk regularly tracks what new releases do and do not deliver.
FAQ
- Is GPT-5.1 bigger than GPT-5? The decimal signals a revision in the same generation; vendors do not disclose parameter counts, so "bigger" is not verifiable — "newer and retuned" is the honest reading.
- What's the difference between Sonnet and Opus? Per Anthropic's documentation, tier: Opus is the largest, most capable line and Sonnet the balanced one, with pricing differences to match.
- Do old models disappear? Via API, often yes — vendors publish deprecation dates and migrate developers to successors. Consumer apps usually switch defaults silently instead.