The History of GPT
From GPT-1 to ChatGPT and GPT-5.6 Sol, Terra, and Luna. How OpenAI's model family became the default assistant.
No single line of models shaped how the public thinks about AI more than OpenAI's GPT family. Its history is also the history of the transformer, RLHF, and reasoning models.
GPT-1 to GPT-3: scaling wins
- GPT-1 (2018) demonstrated that a transformer, pre-trained on unlabeled text and fine-tuned on a task, beat bespoke architectures. The recipe: pre-train, then adapt.
- GPT-2 (2019) scaled the recipe and produced surprisingly coherent text, so much so that OpenAI staged its release over concerns about misuse.
- GPT-3 (2020) made the leap explicit: a 175-billion-parameter model that could perform tasks from a few examples in the prompt, with no fine-tuning at all. "Emergent" capabilities became the headline, and few-shot prompting entered the vocabulary.
The lesson of these three was that intelligence, in this paradigm, is mostly a function of scale. More data and more parameters produced better text, better answers, and better code.
InstructGPT and RLHF
Raw GPT-3 answered prompts but did not follow them well. In 2022 OpenAI introduced InstructGPT, which fine-tuned GPT-3 with RLHF: humans ranked responses, a reward model learned the ranking, and the policy was optimized against it. This alignment step is why ChatGPT, released November 2022, felt like a different product. It was the fastest-growing consumer application in history and started the current era.
GPT-4 and multimodal reasoning
GPT-4 (2023) was the first GPT that handled images as well as text, and its stepwise reasoning made it feel substantially smarter. GPT-4o (2024) added native voice and video, and the o-series introduced deliberate "reasoning tokens" that made chain-of-thought an explicit product feature rather than an emergent quirk.
GPT-5 and the reasoning generation
GPT-5 (2025) consolidated the reasoning improvements into the flagship, with configurable reasoning effort, tool use, and strong coding performance. It was the point where the model became good enough to be a genuinely useful autonomous coder, not just a completion machine.
GPT-5.6: Sol, Terra, Luna (2026)
OpenAI made the GPT-5.6 family generally available in July 2026 across ChatGPT, Codex, and its API. The three tiers share a 1.05M-token context window and a February 2026 knowledge cutoff:
| Model | Positioning | Input / output price |
|---|---|---|
| GPT-5.6 Sol | Frontier model for complex reasoning and coding | $5 / $30 per MTok |
| GPT-5.6 Terra | Balances intelligence and cost | $2.50 / $15 per MTok |
| GPT-5.6 Luna | Cost-sensitive, high-volume workloads | $1 / $6 per MTok |
The tiering is a clear signal of where the market is going: capability calibrated to the job and budget. Use Sol for difficult, high-stakes reasoning; use Terra for routine work; use Luna for high-volume workloads. Tool availability depends on the product surface and plan, so always check the current documentation before designing a workflow around it.
What about GPT-6?
GPT-6 may be next, but OpenAI has not publicly announced a release date or final product details. Treat predictions as predictions, not curriculum facts. The practical lesson is sturdier than any model name: learn the model and harness you have today, then carry your planning, verification, and communication habits forward as the lineup changes.
For your interviews
When a company says "we use GPT-5, Claude, Gemini, and Llama" in a structured interview, they mean a specific model roster. Knowing that GPT-5.6 exists as Sol/Terra/Luna, and roughly which you're using, lets you reason about its strengths and failure modes instead of guessing.