xAI's context windows grew to a million tokens — the price break stayed at 200,000

xAI's published price list carries two sets of rates for every text model: a base rate, and a "long context" rate that applies to requests of 200,000 tokens or more. The higher rate is not a surcharge on the excess — the documentation for Grok 4.6 states that the company charges "different rates for requests which exceed the 200K context window", which reads as the whole request, not the part above the line. The multiplier is the same everywhere: exactly two. Grok 4.6 costs 2.00 USD per million input tokens and 6.00 for output, and 4.00 / 12.00 past the threshold. Grok 4.5 goes from 2.00 / 6.00 to 4.00 / 12.00, Grok 4.3 from 1.25 / 2.50 to 2.50 / 5.00, Grok Build 0.1 from 1.00 / 2.00 to 2.00 / 4.00. Cached input doubles too. What has changed over the past year is the size of the windows, not the position of the break. Grok Build 0.1 holds 256,000 tokens, so 22% of its window falls into the expensive band. Grok 4.5 and Grok 4.6 hold 500,000, so 60% of the window is beyond the line. Grok 4.3 and the Grok 4.20 family hold a million tokens: 80% of the advertised capacity is priced at double. The arithmetic at the boundary is blunt. A 200,000-token prompt to Grok 4.6 costs 0.40 USD at the base input rate. Adding a thousand tokens — half a percent more text — moves the request into the long-context band and, on the plain reading of the documentation, brings the input cost to 0.80 USD. Filling the whole 500,000-token window costs 2.00 USD instead of the 1.00 the headline rate implies. What this does not say: it is not a claim that xAI is expensive. At 1.25 USD per million tokens, Grok 4.3 is one of the cheapest million-token models on the market even at the doubled rate. The point is narrower — the headline price and the headline window describe two different products, and the gap between them widens with every context increase, because the threshold has not moved since the 4.20 generation. xAI has also published a context-compaction feature for developers who need to stay under the line. That is the company's own answer to the same arithmetic.