AI · History · Timeline

Story

Two and a half centuries of artificial intelligence, moment by moment, research, companies and products, people and society, security incidents.

No moment matches your search.

The context race, 2019–2026

In seven years the amount of text a model can hold in view grew roughly ten-thousand-fold, the quiet revolution that made «paste in the whole codebase» possible.

1k10k100k1M10M tokens
GPT-22019
a long email
1,024 tokens
GPT-32020
a short article
2,048 tokens
ChatGPT (GPT-3.5)2022
an essay
4,096 tokens
GPT-42023
a long report
8k → 32k tokens
Claude 2.12023
a 500-page book
200,000 tokens
Gemini 1.5 Pro2024
a small bookshelf
1M (2M preview)
Gemini 2.5 Pro2025
a whole codebase
1 million tokens
Claude Sonnet 4 / 4.52025
five Claude 2.1 windows
1 million tokens
GPT-6 family2026
the Lord of the Rings trilogy
1.05 million tokens

Bars are drawn on a logarithmic scale, at true scale, everything before 2023 would vanish. Every value is a confirmed, shipped context window.

The price collapse

Cost of one million input tokens, at each generation's flagship model.

GPT-3 2020$20.00
GPT-4 2023$10.00
GPT-4o 2024$2.50
DeepSeek V3 2025$0.28
Claude Haiku 5.5 2026$0.10

200× cheaper in six years, intelligence is on its way to becoming a utility.

The compute race

Training compute of frontier runs, in FLOP (floating-point operations).

GPT-2 20191.5×10²¹
GPT-3 20203×10²³
GPT-4 20232×10²⁵
Llama 3 405B 20244×10²⁵
Grok-3-class 2025~5×10²⁶

Log scale: roughly 5–6 orders of magnitude in six years, 4× per year.

Reasoning hits expert level

GPQA Diamond: PhD-level science questions, solved «without tools».

GPT-4 202336%
o1 202478%
GPT-5 202588%
Gemini 3.1 Pro 202694.3%

Above ~70%, the benchmark stops discriminating: it is saturated.