A million tokens of room to think: what big context is actually for

Both models in the Raven family, Flash and Max, carry a one-million-token context window and 128,000 tokens of output. That first number has become a bragging-right in the industry, which is unfortunate, because it is also one of the most practically useful specs a model has. This post is about what a million tokens actually holds, and about what a workday feels like once the model stops forgetting how it started.

The units, made concrete

A million tokens is about 750,000 words of English, at roughly three-quarters of a word per token. That is five to seven full novels, a complete contract portfolio with every schedule, most mid-sized codebases with dependencies, and a year of a team's documentation, assuming the team writes like humans rather than lawyers.

On the way out, 128,000 tokens covers a full book chapter, or a whole grant application, or a day's worth of code, in one piece rather than in fragments you stitch together.

The forgotten-start problem

Every long conversation with a conventional window ends the same way. Somewhere around the middle of a real working session, the earliest material silently falls off the edge. The model keeps answering, confidently, about a document it can no longer see. Summaries get re-summarized until the detail you needed is dead, and you end up re-pasting key passages "just in case", which is you doing the model's remembering for it.

A million tokens removes that cliff for almost everything real work throws at a model. Paste the spec at 9 am and it is still fully there at 5 pm, with the 5 pm answer actually using it.

What the room is for

Three kinds of work feel the difference fastest. In DCode, whole-repo reasoning stops being a metaphor: the agent holds the dependency graph and every caller of every function at once, which is precisely what "refactor the checkout flow" requires. Document work changes shape too. Drop in a 300-page agreement with its appendix and last year's version, and ask what changed. The model reads everything; summaries never enter into it. Research may benefit most of all, since the entire collected source set lives in one context and the synthesis cites what the sources said rather than what a compressed digest remembered.

Vision rides along too. Screenshots, scans, and charts share that same context with the text. Paste a dashboard and it joins the analysis; nothing sits unopened in an attachment pile.

Numbers, if you like numbers

Context does not trade off against speed here. Raven Flash holds the million-token window while keeping sub-second first tokens, which is why Flash carries both the long window and the velocity tier lead on Terminal Bench 4.0, and why Raven Max pairs the same window with the top score on long-horizon agentic benchmarks like Agent's Last Exam. The long window was not bolted on after the benchmarks; the benchmarks are partly a measure of it. A model holding the whole problem behaves differently from one holding a summary of the problem, and the eval scores follow.

You can feel the difference yourself without a benchmark: open the playground, paste the longest real document you have lying around, and ask about the part you know is buried near the end.