Context Window Comparison

Context windows are the most-quoted and least-understood model spec. This table puts every major model side by side — total window, maximum output per response, current pricing and streaming speed — sortable by any column, with the data verification date shown (2026-06-11).

Two distinctions trip people up. First, window versus max output: a 200k window does not mean 200k of generation — output caps are typically 8-128k, and input plus output must fit the window together. Second, advertised versus usable context: published needle-in-a-haystack benchmarks show retrieval accuracy degrading well before windows fill, especially for facts buried mid-context. A million-token window is a real capability, but treating it as a database is how agents start hallucinating file contents they read an hour ago.

The cost dimension is on the same table because window and price interact: carrying a big context is not free, it is re-billed every call. At Claude Sonnet rates, holding 150k tokens of session history costs about $0.45 of input per turn before the model writes a word — caching reduces it, compaction eliminates it. The bars next to each window size keep the scale honest; a 1M window is five times Claude's, and the bar shows it.

Speed figures are median public benchmark numbers for streaming throughput and vary with load, region and prompt shape — treat them as ratios between models rather than guarantees. For how your own sessions actually fill these windows, FORG tracks context size per turn across every real session.

When two models look equivalent on this table, the tiebreakers usually live elsewhere: tokenizer efficiency (the same document can differ ten percent in token count between vocabularies), cache pricing, and how gracefully each model degrades as its window fills. The comparison here is the factual baseline; pair it with the Context Rot Simulator for the degradation story and the Model Capability Picker when the decision is about task fit rather than specs.