AI in Business
Long-Context Surcharge in plain English.
Also known as: long-context pricing,context tier pricing,200K surcharge,272K surcharge
The one-sentence version
A higher per-token rate that applies once a prompt crosses a size threshold, usually charged on the whole request rather than only the tokens over the line.
Most 2026 frontier models advertise a 1M-token context, but the price per token is not flat across it. Above a threshold the vendor charges more, and in most cases the higher rate applies to every token in that request, output included. The thresholds and multipliers differ: OpenAI's GPT-6 Astra and GPT-6.1 Sol double input and raise output 1.5× above 272K input tokens; xAI's Grok 4.7 doubles everything at 200K; Anthropic's Claude Haiku 5.5 moves from $0.10/$0.50 to $0.50/$2.50 above 100K, while Sonnet 5.5 has no long-context premium at all. Google's Gemini 4 Argon has a single introductory rate with a 1M output limit. The practical effect is a cliff: a 280K-token request to Sol costs more than twice a 270K one. Teams that stuff whole codebases into prompts should measure where their requests land, trim or chunk to stay under the threshold, lean on caching (cached input is often discounted even above the line), and compare models on their real prompt sizes rather than on headline rates. Cost calculators, including ours, usually show the base rate; the surcharge lives in the fine print.