The Context Window Grew 488x. The Search Box Did Not
The median AI model you can buy today has a context window of one million tokens. In 2022 the working figure was around two thousand. That is a 488-fold increase in four years.
Over the same period, the interface everyone spent twenty years optimising for did not change at all. A text field. About eight words.
That gap explains more about why people stopped searching for certain things than any argument about search quality does, and this piece walks through why.
What a context window is, in plain terms
A context window is everything the model can look at while it writes its answer. Anthropic's own documentation describes it as the model's working memory – the conversation so far, whatever you have attached, and the reply it is generating, all counted together in tokens.
A token is roughly three quarters of a word.
So at two thousand tokens, you can ask a question. At a million, you can hand over the problem: a codebase, a year of email, a contract with its schedules, a folder of PDFs, a recording of a meeting.
The size of the thing you can put in front of a model grew by nearly three orders of magnitude. That is the asymmetry. Not "AI answers better than Google". The unit of input changed.
The plain text model stopped being made
Alongside the window, the shape of what ships changed completely.
In 2023 this field released 25 models whose category was simply "text". In the 250 days of 2026 so far it has released zero. Not fewer. None.
What replaced them:
- 60.5% of this year's releases accept more than text: images, PDFs, audio, video.
- 74.1% name reasoning as a capability. In 2023 none did. The first appeared in 2024; there are now 110.
- A third of all releases carry a compound category such as "multimodal / reasoning", which is why capability counts overlap rather than adding up to the release total.
Put the window and the modalities together and the real change appears. A text-only model with a small window could only answer a question someone had already written down. A model that accepts a screenshot, a PDF, a voice note or a video answers questions people never bothered to type, because typing them was the hard part.
Every search interface ever built assumes you can express your question in words, in a box. You cannot photograph a problem into Google and get an answer worth having. You can into an assistant.
It got cheaper while it got bigger
The usual assumption is that more capability costs more. At the low end, the opposite happened.
The cheapest capable model fell from $0.60 to $0.05 per million output tokens between 2023 and 2026 – twelve-fold. Frontier pricing rose 25%.
That divergence is the commercially important line in the whole release history. Cheap intelligence is why an assistant now sits inside products where nobody would have paid for one, which is how assistants ended up in the tools people already had rather than at a destination they had to visit. Expensive intelligence is where the frontier labs earn their margin.
Two markets moving in opposite directions, almost always reported as one.
The consequence you can measure
None of this would matter commercially if search demand had held. It did not, and it did not fall evenly.
I sorted six keywords into two groups before pulling any trend data. Group A: questions an assistant can finish inside the tool you are already in. Group B: physical, local and comparison questions, where the answer is not the deliverable.
Group A fell 64.2% from peak. Group B fell 4.5%. A 14.2-fold difference, same index, same window, same rule.
The steepest single fall was 74.7%, and no competitor was involved. Nobody out-ranked that page. The question stopped being asked, because what it was asking for now fits inside a context window next to the user's actual code.
Six keywords describe a mechanism, not a market. The mechanism is what transfers: as the window grows and the modalities widen, the set of questions that must be typed into a box keeps shrinking.
Who is paying for all of this
Worth knowing before you plan around today's prices, because cheap inference is the load-bearing assumption in most AI product plans.
Disclosed funding across the private labs runs to roughly $356bn. The funding-to-booked-revenue ratio is about 12.4x, and only 7 of 15 labs have an audited revenue figure. Every headline marked ARR or run rate is one month multiplied by twelve.
That does not make the technology a bubble, and I am not claiming it does. It means something narrower: the $0.05 floor and the release cadence behind it are currently paid for by capital rather than income. If capital tightens, the free tier goes first and the price floor second.
Two other changes are worth having on your risk register. Seven models on the current shelf carry a category naming cyber capability, and the first model graded critical on cyber capability by its own maker shipped this year. And OpenAI has confirmed its agents editing third-party sites unprompted – the open web is no longer only something assistants read.
What a bigger context window means for your plans
The context window is a technical number with three commercial consequences, and none of them are on most roadmaps.
Re-check the boundary every two quarters. A query that was safe from substitution in 2024 – anything needing an image, a document, a diagram or a screenshot – may not be safe now. The set of tasks an assistant can finish grows with every release that names reasoning, and 74.1% of this year's do.
Think about the input, not the phrasing. If your customers' real question is "here is my situation, what should I do", it was never well served by a search box, and it never showed up in a keyword tool because nobody could type it. That demand did not appear in 2026. It became askable in 2026.
Recompute anything you shelved on cost. If a feature was uneconomic at $0.60 per million tokens, run the numbers again at $0.05 before shelving it permanently. A lot of abandoned 2023 roadmap items are viable now purely on price.
And plan for a visitor who is not a person. When an agent browses instead of a human, dwell time and engagement signals collapse, which breaks measurement before it breaks the funnel.
Questions people ask about the context window
These are the questions searched most often alongside context window, answered using the release data above.
What is a context window in AI? Everything the model can see while producing an answer: your prompt, the conversation so far, any attached files, and the reply itself. It is measured in tokens, where a token is roughly three quarters of a word.
What is a context window in an LLM, in practical terms? It is the size of the problem you can hand over in one go. At two thousand tokens you ask a question. At a million you can attach a codebase or a year of documents and ask about all of it at once.
How big are context windows in 2026? The median model currently on sale holds one million tokens. Several vendors advertise larger extended modes. In 2022 the working figure was about two thousand.
Does a bigger context window make a model better? Not on its own. It changes what you can ask, not how well the model reasons. The releases that matter most combine a large window with reasoning and multimodal input, which is now the majority of what ships.
Why does the context window matter for search and SEO? Because it changes what people can ask without typing. Questions that need a document, an image or a screenshot were never searchable, so they never appeared in keyword volume. They are askable now, and that demand was never in your keyword tool to begin with.
Context, capability, pricing and status figures: a workbook of 261 model releases across 15 labs, 11 June 2018 to 3 September 2026, captured 7 September 2026. Context window figures take the first integer in each cell and honour only the suffix attached to it, which is conservative and can understate a vendor's headline claim. Cohort figures: Ahrefs monthly volume history, groups assigned before trend data was pulled. Full method in the report, How AI Changed the Way People Search.
Tags
