Context length turns a low-cost model into a workload-design question
Thesis: Cursor’s listing of Claude Haiku 5.5 makes context length a material part of the model’s cost decision for agent workflows. Mechanism: Cursor prices requests above 100,000 input tokens at five times its standard rates, so long-context sessions face a distinct per-token cost tier even though the model supports context windows up to 1 million tokens. Why now: Cursor has added the model to Settings → Models and published both its rates and its own benchmark figures, giving users actionable terms to evaluate.
Observed facts: Cursor’s documentation lists rates of $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 input tokens. It lists $0.50 per million input tokens and $2.50 per million output tokens above that threshold. The page reports a 48.4% CursorBench score at maximum effort and identifies 1 million tokens as the maximum context window. First implication: Teams using Haiku 5.5 in long agent sessions should account for the higher tier rather than extrapolating the short-request price. Evidence boundary: These prices and the CursorBench result are claims in Cursor’s documentation; the source does not independently establish comparative cost or performance in production workloads.
First watch condition: Watch for changes to Cursor’s published 100,000-token threshold, rate schedule, or model availability in the model picker. The evidence establishes what Cursor documents, not how often actual sessions cross the threshold or what their end-to-end costs will be.
This starts a distinct analytical record because no prior New Runtime record was supplied; no connection to other records is asserted.