This is a historical article. It traces how AI language models developed from 2017 up to the end of 2025, and it is deliberately written in the past tense. Model names, versions, context windows, and pricing change every few weeks — nothing here should be read as advice about which model to use today. For current models, always go to the vendor's own documentation, linked at the end of this article. Last reviewed: July 2026.
Why look backwards at all?
Artificial intelligence is one of the fastest-moving fields in technology, and that speed has a side effect: almost everything written about "the best model" is wrong within months. What does not go stale is the shape of the story — which ideas actually moved the field forward, and why.
That is what this article is for. Instead of comparing today's models, it walks through the generations that got us here, what each one changed, and which of those changes turned out to be permanent. If you are trying to understand why modern models behave the way they do, this history is more useful than any benchmark table.
The Transformer moment (2017)
The term "artificial intelligence" dates back to the 1950s, but the architecture behind every modern language model arrived in a single paper.
In 2017, a team of Google researchers published "Attention Is All You Need", introducing the Transformer architecture. Its key idea was the attention mechanism: instead of processing text strictly in sequence, the model learns which parts of the input matter most for each part of the output, and it can do that for the whole input in parallel.
Two consequences followed, and both mattered enormously:
- Training could be parallelized, which meant it could scale with hardware in a way earlier recurrent architectures could not.
- Long-range context became tractable, so models could relate words far apart in a document.
Every model family in this article — BERT, GPT, Claude, Gemini, Llama — is built on that foundation. If you read one paper from this era, read that one.
2018–2019: BERT, GPT-2, and the scaling hypothesis
2018 brought two developments that split the field into two useful directions.
Google's BERT was designed to understand text. It read input in both directions at once, which made it excellent for classification, search ranking, and question answering. It was quietly one of the most commercially deployed models of its era — it went into Google Search.
OpenAI's GPT series went the other way: it was designed to generate text, predicting one token at a time. That framing seemed more limited at first, and turned out to be far more general.
Then in 2019, GPT-2 demonstrated something uncomfortable and important: making the same architecture bigger, with more data, produced better results across tasks nobody had specifically trained it for. This became known as the scaling hypothesis, and it set the industry's direction for the next five years.
2020–2022: GPT-3 and the ChatGPT moment
GPT-3 (2020) was the point where scale stopped being an academic curiosity. At 175 billion parameters, it could perform tasks from a plain description and a couple of examples in the prompt — no retraining required. That capability, in-context learning, is the reason prompting became a skill at all.
But GPT-3 was an API for developers. Most people never touched it.
The change came in November 2022 with ChatGPT — a chat interface built on GPT-3.5. The underlying capability was not radically new; the packaging was. Putting a text box in front of the model took AI from a research topic to something hundreds of millions of people used for writing, coding, learning languages, and everyday problem-solving. It remains the clearest example in recent tech history of interface mattering as much as capability.
Two techniques from this period are worth knowing because they still shape model behaviour:
- Instruction tuning — training the model to follow instructions rather than merely continue text.
- RLHF (Reinforcement Learning from Human Feedback) — using human preference ratings to shape which answers the model favours. This is why models are helpful and polite by default, and also why they sometimes agree with you too readily.
2023–2025: the model boom
After ChatGPT, the field went from a handful of serious labs to a crowded market. This is the period where the "which model should I use?" question became genuinely hard — and it is also the period whose specifics have aged the fastest.
OpenAI: GPT and the reasoning models
OpenAI spent this period doing two things in parallel.
The GPT line became multimodal — first images, later audio and video — and grew steadily larger context windows. These were the generalists: fluent, fast, good at writing and conversation.
The "o" series (o1, o3, o4-mini) was the more interesting departure. These were reasoning models: instead of answering immediately, they generated an internal chain of thought, explored alternatives, and checked their own work before responding. They were slower and more expensive, and substantially better at mathematics, programming, and multi-step logic.
That split — fast generalist versus deliberate reasoner — was the defining architectural debate of the era. It has since been resolved in an interesting way: rather than shipping two separate model lines, vendors folded reasoning into single models that decide how much to think based on the difficulty of the request.
Google Gemini: multimodality and context length
Gemini launched in December 2023, and Google leaned on two advantages.
The first was native multimodality. Where GPT had text models that later gained vision, Gemini was trained from the start on text, images, audio, and video together — so combining them in a single analysis was natural rather than bolted on.
The second was context window size. Gemini pushed to roughly a million tokens well before competitors, which made "paste the entire codebase" or "analyse this whole book" a realistic workflow instead of an exercise in chunking.
Google also popularised Mixture-of-Experts (MoE) in mainstream models: rather than activating the whole network for every token, MoE routes each request through a subset of specialised sub-networks. The practical effect is a model with a large total parameter count that costs much less to run than that count suggests. MoE is now near-universal at the frontier.
A period artifact: the LMArena leaderboard as it looked in 2025. Every model in this screenshot has since been superseded — which is precisely the point about benchmark snapshots.
Anthropic Claude: safety as an architecture choice
Anthropic was founded in 2021 by former OpenAI researchers, and its models arrived in 2023 with a distinct training approach: Constitutional AI.
Rather than relying only on human preference ratings, Claude was trained against an explicit written set of principles — a "constitution" — that it used to critique and revise its own outputs. The goal was to reduce harmful or dishonest answers without needing a human to label every possible failure.
The Claude line settled into a three-tier naming scheme that has proven durable: Haiku (fast and cheap), Sonnet (the balanced workhorse), and Opus (the most capable). Through 2024–2025 Claude built a specific reputation among developers for code generation and for following long, detailed instructions carefully — which is how it ended up embedded in a lot of developer tooling.
Open-source models: the pressure from below
The most underrated development of this period was open-weight models becoming genuinely competitive.
- Meta's Llama made capable open weights normal, and seeded an enormous ecosystem of fine-tunes and local-inference tooling.
- Mistral AI established that a European lab could compete at the frontier, with an emphasis on small models that punched above their weight.
- DeepSeek was the shock of the period: a Chinese lab producing frontier-adjacent reasoning models at a fraction of the expected training cost, released under a permissive licence.
- Qwen (Alibaba) shipped broad, multilingual, permissively licensed families that became a default starting point for a lot of self-hosted work.
The strategic effect mattered more than any single model. Open weights put a ceiling on what closed vendors could charge for mid-tier capability, and made self-hosting a real option for teams with privacy, cost, or latency constraints.
What actually changed between generations
Strip away the version numbers and the same handful of shifts explain most of the progress:
| Shift | What it changed |
|---|---|
| Transformer + attention (2017) | Made training parallelizable and long context tractable — the foundation for everything after |
| Scaling (2019–2020) | Bigger models on more data got better at tasks they were never trained for |
| Instruction tuning + RLHF (2022) | Turned text predictors into assistants that follow instructions |
| Chat interfaces (2022) | Took AI from an API to a mass-market product |
| Multimodality (2023–2024) | One model handling text, images, audio, and video instead of separate specialised models |
| Long context (2024) | Whole repositories and books as input, rather than carefully chunked excerpts |
| Mixture-of-Experts (2024–2025) | Large total capacity at a fraction of the inference cost |
| Reasoning / thinking models (2024–2025) | Models that spend compute deliberating before answering — a large gain on maths, code, and logic |
| Tool use and agents (2025) | Models that call external tools and act, instead of only producing text |
Those are the durable lessons. Notice that none of them is a model name.
Where things stand now (July 2026)
Briefly, and with the explicit warning that this section will be out of date before most other parts of this article:
- OpenAI ships the GPT-5.6 family, split into capability tiers rather than a separate reasoning line. The older "o" series is being retired.
- Google ships Gemini 3.x, with Gemini 4 reported to be in pre-training.
- Anthropic ships the Claude 5 family — Opus, Sonnet, and Haiku tiers, with reasoning built in rather than exposed as a separate model.
- Open-weight models from Qwen, DeepSeek, GLM, Llama, and Mistral remain competitive for a wide range of work, increasingly under permissive MIT or Apache-2.0 licences.
The clearest structural trend since 2025 is that the "reasoning model versus chat model" distinction has collapsed. Modern models decide internally how much deliberation a request deserves, which removes a choice developers used to have to make by hand.
How to use a history like this
If you came here looking for a model recommendation, the honest answer is that no article can give you one that stays correct. What you can do instead:
- Check the vendor's own model documentation. It is the only source updated on the same day as the models. All three are linked below.
- Compare on your own task, not on a leaderboard. Public benchmarks are a weak proxy for whether a model handles your prompts, your domain, and your edge cases.
- Assume the tier names outlast the version numbers. "Fast and cheap / balanced / most capable" has been a stable shape across vendors for years, even as the models behind those labels change constantly.
- Watch for the shifts, not the releases. The table above took nine years to fill in. Individual model launches matter much less than the structural changes they occasionally represent.
The models in this article are all obsolete. The reasons they mattered are not.