OpenAI’s catalogue lists over a hundred models. Four of them are the answer. Everything else is a previous generation, a specialised variant, or something already scheduled for shutdown. Here is the short list, and how to read the long one.OpenAI, Models, read at source 16 Sep 2026. The page separates “Featured models” from the full catalogue; the featured list has four entries.
- Astra, Sol, Terra, Luna. Those are the four current models.
- Astra is the ceiling — “our most capable model, built for the hardest end-to-end work”.
- Terra is the balance and Luna is the cheap one.
- Everything named GPT-4 or o-something is previous generation.
- Anything with
previewin the name can go in two weeks.
Every price here was correct when this site last checked it, between 16 Sep and 22 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: OpenAI pricing.
The four that matter
| Model | What OpenAI says it is for | In / out |
|---|---|---|
| GPT-6 Astra | “Our most capable model, built for the hardest end-to-end work” | $10 / $50 |
| GPT-5.6 Sol | “Flagship model for complex professional work” | $4 / $20 |
| GPT-5.6 Terra | “GPT-5.6 model that balances intelligence and cost” | $2 / $12 |
| GPT-5.6 Luna | “GPT-5.6 model optimized for cost-sensitive workloads” | $0.20 / $1.20 |
OpenAI, Models and Pricing, both read at source 16 Sep 2026. Prices are per million tokens, standard tier, short context.
Fifty times separates Astra from Luna on input. The four are a ladder, not a menu — the question is which rung each part of your workload needs, not which model is best.
What they all do
OpenAI states one thing about the whole current family, which saves checking each model page: “All latest OpenAI models support text and image input, text output, multilingual capabilities, and vision.”
So the choice between the four is cost and capability depth, not feature availability. What is not uniform is price behaviour — long context roughly doubles every rate, and the detail is on what the OpenAI API actually costs.
How to read the other hundred
The full catalogue is mostly history and specialism. Four rough groups:
- Previous generations. GPT-5 through 5.5, GPT-4 and its variants, and the entire o-series. OpenAI marks these plainly — o3 is “succeeded by GPT-5”, and o4-mini is “succeeded by GPT-5 Mini”.
- Task-specific variants. Codex models for agentic coding, Realtime and Audio families for voice, Transcribe for speech-to-text, embeddings and moderation. The newest voice model, GPT-Live 1, became generally available on 10 September 2026 at $0.05 a minute for the voice session, with any backend model billed on top.OpenAI, API changelog, 10 Sep 2026, read at source 22 Sep 2026: “GPT-Live 1 is now generally available in the API.” and “Voice sessions cost $0.05 per minute, billed per second; backend model and tool usage is charged separately.”
- Gated specialist models. GPT-5.6 Cyber is described as “our most advanced cybersecurity model for authorized vulnerability research and security testing”, and GPT-Rosalind as “life sciences reasoning for approved organizations”. Both are access-controlled rather than open to any account. GPT-Rosalind became generally available to approved organisations on 8 September 2026, with billing from 5 October.OpenAI, API changelog, 8 Sep 2026, read at source 22 Sep 2026: “GPT-Rosalind (
gpt-rosalind-research) is now generally available through the trusted-access program for approved internal life sciences research.” and “Billing begins on October 5, 2026.” - Open weights.
gpt-oss-120bis described as the “most powerful open-weight model, fits into an H100 GPU”, with a 20b sibling.
The word that should stop you
If a model name contains preview, do not build a product on it. OpenAI is direct about the risk: preview models “may be retired with much shorter notice, such as 2 weeks”, and it says it does not recommend them “for business-critical production workloads unless you can migrate on short notice”.OpenAI, Deprecations, read at source 16 Sep 2026.
Several models still listed in the catalogue already have shutdown dates attached. The notice periods, and what a retirement actually does to your calls, are on when AI models retire.
Against the others
Anthropic takes the opposite approach: four models, no back catalogue on the same page, and a stated default. That comparison is on which Claude model to use, and the prices sit side by side on the four-vendor table.
What this page could not verify
- Any performance claim. Every description is OpenAI’s own.
- Context window sizes. Held on individual model pages, not read here.
- Which model suits your task. That needs your evaluations, not a catalogue.
- Access requirements for the gated models. Described as approved or authorised without the process being stated.
OpenAI, Models, Pricing and Deprecations. All read at source on 16 September 2026.
Model lineups change monthly. Re-read before committing.