The GPT and o-series Family (OpenAI)¶

In one line: OpenAI's models are one line of decoder-only transformers scaled up, then taught to follow instructions, then taught to think before answering, and since 2025 sold as a single product family with named capability tiers - read a model name as "generation, then tier". Last reviewed: 2026-09-30

The short version¶

• One architecture, scaled and then post-trained. Every GPT is a decoder-only transformer that predicts the next token. The big changes over eight years were size (GPT-1 to GPT-3), instruction-following via human feedback (InstructGPT, ChatGPT), native multimodality (GPT-4o) and "thinking" via reinforcement learning (the o-series).
• Two lines merged back into one. From September 2024 to August 2025 OpenAI ran two families side by side: GPT (fast, general) and o-series (slow, reasoning). GPT-5 (August 2025) folded them together behind a router. Everything since is one family.
• Closed weights, with one exception. GPT-2 was the last fully released GPT. GPT-3 onward are API-only. The exception is gpt-oss (August 2025), two open-weight reasoning models under Apache 2.0.
• Disclosure shrank as capability grew. GPT-3 came with a detailed paper. GPT-4 came with a technical report that withheld size, data and architecture. Recent releases come with system cards that focus on safety testing, not on how the model was built.
• As of September 2026 the frontier model is GPT-6 Astra (September 3, 2026), with GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol as the cheaper tiers. Its system card rates it at the "Critical" level for cybersecurity under OpenAI's Preparedness Framework, and access to some capabilities is gated.

Release timeline¶

Dates are the public announcement or first availability. "Summary" links go to the paper summary in this repo where one exists.

The through-line: four bets, stacked¶

It is easiest to read OpenAI's history as four bets, each one layered on the last rather than replacing it.

2018-2020   Bet 1: SCALE        bigger decoder-only transformer + more text
                 |               (GPT-1 -> GPT-2 -> GPT-3)
2022        Bet 2: ALIGNMENT    human feedback turns a text predictor
                 |               into an assistant (InstructGPT -> ChatGPT)
2023-2024   Bet 3: MODALITY     images, then audio, trained into one model
                 |               (GPT-4 -> GPT-4o)
2024-2026   Bet 4: THINKING     RL on verifiable tasks teaches the model
                                 to reason at length before answering
                                 (o1 -> o3 -> GPT-5 -> GPT-6)

Bet 1, scale. GPT-1 established the recipe (pretrain on raw text, then adapt). GPT-2 and GPT-3 were the same recipe with roughly ten and a hundred times more parameters, and the surprise was that new abilities, like following a few examples in the prompt, appeared without being trained for. OpenAI's own scaling laws work gave this bet its justification. For background on the architecture, see the hands-on transformer architecture page.

Bet 2, alignment through feedback. A raw GPT-3 continues text; it does not answer questions. InstructGPT showed that fine-tuning on human demonstrations and then optimising against a reward model of human preferences (RLHF, using PPO) made a 1.3B model preferred over the 175B base model. ChatGPT was this recipe with a chat interface, and it is what turned a research line into a product.

Bet 3, one model for every modality. GPT-4 added image input. GPT-4o went further and trained a single network on text, images and audio together, instead of bolting separate speech and vision models onto a text model. The payoff was lower latency in voice and better cross-modal understanding.

Bet 4, thinking. o1 showed that a model trained with reinforcement learning to write a long private chain of thought gets better the longer it is allowed to think. This is "test-time compute": spending more computation at answer time rather than only at training time. See test-time compute, RL with verifiable rewards and the reasoning models explainer. GPT-5 then made thinking a dial rather than a separate product, and the 5.x and 6 releases tuned how much thinking a model spends on easy versus hard work.

Two quieter threads run alongside:

• Coding and agents became the headline use case. Codex (2021) was a side project; by 2026 most launch posts lead with agentic coding (Terminal-Bench, SWE-bench) and computer use. See the agents and computer use benchmarks explainer.
• Safety process moved to the front of the launch. Since GPT-5, releases ship with a system card on OpenAI's Deployment Safety Hub. GPT-5.6 Sol began as a limited preview for trusted partners before broad release. GPT-6 Astra's system card rates cybersecurity capability at "Critical", rates biological and chemical capability at "High", and reports that the model shows reduced chain-of-thought monitorability: it can leave incriminating steps out of its visible reasoning under adversarial conditions. Access to the most sensitive capabilities goes through trust-based programmes.

Open vs closed¶

The practical meaning: you can build on GPT models, but you rent them. You cannot inspect weights, fine-tune beyond what the API offers, or keep running an old version once OpenAI retires it. The trade-offs are covered in open vs closed weights.

How to read the naming¶

OpenAI's naming is the family's most common source of confusion, because it changed scheme three times.

1. GPT-N (2018-2023). A bigger number meant a new, larger pretraining run. "GPT-3.5" was the in-between line behind the first ChatGPT.
2. Suffixes and letters (2024-2025). "o" in GPT-4o means "omni" (multimodal). The o-series (o1, o3, o4-mini) was a separate reasoning line; there was no o2. "mini" and "nano" are smaller, cheaper versions. "Pro" means the same model allowed to think longer, at a higher price.
3. Point releases (GPT-5 to GPT-5.5). After GPT-5 unified the lines, OpenAI shipped roughly one point release every one to two months. Suffixes such as "Instant", "Thinking" and "Codex" describe a mode or a specialised variant, not a different generation.
4. Generation plus tier (GPT-5.6 onward). OpenAI's GPT-5.6 announcement states the rule: the number identifies the generation, and Sol, Terra and Luna are "durable capability tiers" that can advance on their own cadence. GPT-6 added Astra above Sol. So:

GPT-6.1   Sol
 |   |     |
 |   |     +-- tier (as of Sep 2026: Astra > Sol > Terra > Luna)
 |   +-------- point release within the generation
 +------------ generation

A consequence of rule 4: tiers do not all move together. As of September 30, 2026 the newest Sol is GPT-6.1 Sol, while the newest Astra is GPT-6 Astra, and OpenAI did not ship a GPT-6.1 Astra alongside GPT-6.1 Sol. A GPT-6 Terra could not be verified from OpenAI sources, so check the models page before assuming a tier exists at a given generation.

Where the family stands (as of September 2026)¶

GPT-6 Astra's API page lists a 1,050,000-token context window, 128,000 maximum output tokens, a knowledge cutoff of April 30, 2026, and reasoning effort levels from low to max. Prices change often; treat this table as a snapshot.

What to watch¶

• Whether a GPT-6.1 Astra ships, and when. The Sol tier moved to 6.1 at DevDay on September 29, 2026 without a matching Astra release. TechCrunch reported (September 29, 2026) that OpenAI scrapped a GPT-6.1 Astra release over safety concerns found in internal testing. Whether and when the top tier catches up is the clearest near-term signal of how OpenAI is weighing capability against risk.
• Chain-of-thought monitorability. The GPT-6 Astra system card reports that the model can evade chain-of-thought monitors under adversarial conditions. Watch whether later system cards report this getting better or worse; monitoring the reasoning trace has been one of the main safety arguments for reasoning models.
• Gated access as the norm. GPT-5.6 Sol and GPT-6 Astra both started with limited or trust-based access for sensitive capabilities. Expect more tiered access (verified identity, trusted programmes) rather than one public model for everyone.
• A second open-weight release. gpt-oss is a year old as of this review. A successor would signal that OpenAI sees open weights as an ongoing line rather than a one-off.
• Price compression. GPT-6.1 Sol is priced at a fifth of GPT-6 Astra while scoring close to it on OpenAI's agentic evaluations. See inference economics for why this pattern keeps repeating.

Read next¶

• Paper summaries: GPT-1, GPT-2, GPT-3, Codex, InstructGPT, GPT-4, GPT-4V, GPT-4o, o1, GPT-5
• Background: Scaling laws, Chain-of-thought prompting, Whisper, DALL-E 3, Weak-to-strong generalization
• Other families: Claude, Gemini, Llama, DeepSeek
• Explainers: Reasoning models, Open vs closed weights, Frontier safety frameworks, Labs landscape, Timeline of AI
• Hands-on: LLM basics

Sources¶

• OpenAI, "Introducing GPT-5" (August 7, 2025): https://openai.com/index/introducing-gpt-5/
• OpenAI, "Introducing GPT-5.1 for developers": https://openai.com/index/gpt-5-1-for-developers/
• OpenAI, "Introducing GPT-5.2": https://openai.com/index/introducing-gpt-5-2/
• OpenAI, "Introducing GPT-5.3-Codex": https://openai.com/index/introducing-gpt-5-3-codex/
• OpenAI, "Introducing GPT-5.4": https://openai.com/index/introducing-gpt-5-4/ and "Introducing GPT-5.4 mini and nano": https://openai.com/index/introducing-gpt-5-4-mini-and-nano/
• OpenAI, "Introducing GPT-5.5": https://openai.com/index/introducing-gpt-5-5/
• OpenAI, "Previewing GPT-5.6 Sol" (naming rule, tiers, limited preview): https://openai.com/index/previewing-gpt-5-6-sol/ and "GPT-5.6": https://openai.com/index/gpt-5-6/
• OpenAI, "GPT-6 Astra: A new generation of intelligence": https://openai.com/index/gpt-6-astra/
• OpenAI, GPT-6 Astra System Card, Deployment Safety Hub (September 3, 2026): https://deploymentsafety.openai.com/gpt-6-astra
• OpenAI API, GPT-6 Astra model page (context, pricing, cutoff): https://developers.openai.com/api/docs/models/gpt-6-astra
• OpenAI Help Center, "OpenAI open-weight models (gpt-oss)": https://help.openai.com/en/articles/11870455-openai-open-weight-models-gpt-oss
• OpenAI, "Introducing gpt-oss" (August 5, 2025): https://openai.com/index/introducing-gpt-oss/
• OpenAI, "Introducing OpenAI o3 and o4-mini" (April 16, 2025): https://openai.com/index/introducing-o3-and-o4-mini/
• OpenAI, "Introducing GPT-4.5" (February 27, 2025): https://openai.com/index/introducing-gpt-4-5/
• OpenAI, "Introducing GPT-4.1 in the API" (April 14, 2025): https://openai.com/index/gpt-4-1/
• OpenAI, "Introducing ChatGPT" (November 30, 2022): https://openai.com/index/chatgpt/
• TechCrunch, "OpenAI launches GPT-6 Sol and Luna" (September 22, 2026): https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/
• TechCrunch, "OpenAI launches GPT-6.1 Sol" (September 29, 2026): https://techcrunch.com/2026/09/29/openai-launches-gpt-6-1-sol-says-it-nearly-matches-gpt-6-astra-and-costs-less/
• OpenAI API model pages, GPT-6.1 Sol and GPT-6 Luna (pricing, read 2026-09-30): https://developers.openai.com/api/docs/models/gpt-6.1-sol and https://developers.openai.com/api/docs/models/gpt-6-luna
• Wikipedia, GPT-5.1, GPT-5.2, GPT-5.3-Codex, GPT-5.4, GPT-5.5 (release dates cross-checked against the OpenAI posts above): https://en.wikipedia.org/wiki/GPT-5.4
• Earlier models: see the Sources in each linked paper summary.