How much a model can read at once (its context window) and how much each token costs have moved in opposite directions at a staggering pace. Context windows grew roughly 500× in three years; the price of frontier-grade intelligence fell by orders of magnitude. This tool charts both curves on a real model lineage, marks the milestones that mattered, and shows the controls teams use to keep spend in check.
Each dot is a model release. The y-axis is the context window — how many tokens (roughly ¾ of a word each) the model can hold in mind at once. Switch to a log scale to see the steady multiplicative climb; tap any point for the details.
From a 200K-token window in late 2023 to a full million today — enough to read an entire code repository, a stack of contracts, or a short book in a single prompt.
Progress didn't arrive evenly. A handful of thresholds — in context, in price, and in how tokens are billed — each unlocked a new class of use case. These are the ones that moved the needle for the teams we work with.
List price per million tokens at launch, plotted over time on a log scale. The long trend is steeply down — the same money buys far more capability every year. The exception is the very top of the range: flagship models price back up because they do more per token. Toggle input vs. output pricing, and note that output always costs several times more than input.
A small, fast model now answers for cents per million tokens — work that cost dollars a couple of years ago. Cheaper tokens are what turned one-off demos into always-on products.
When each token is cheap and the window is huge, it's easy to spend a lot without noticing. These are the levers — built into the API and the platform — that teams use to cap consumption, bound cost, and keep an agent from running away with the budget.
Falling token prices and million-token windows change which problems are worth solving with AI — and which architectures actually pay off. That's the work we do with the teams we advise: turning a trend line into a roadmap, a cost model, and a build that holds up in production.
Talk to us about your AI roadmap