Context Graph Makes Haiku 5.5 23% faster, for 19% less money, in fewer steps

A benchmark of Claude Haiku 5.5 answering 17 engineering questions across a public GitHub organization with 802 repositories found that providing a single upfront Context Graph API query reduced average end-to-end response time by 23%, model cost by 19%, input-token use by 20%, and tool calls by 9% across 102 graded runs. The Context Graph acts as an index of services, endpoints, schemas, and cross-repository relationships, giving the model candidate locations to verify in code rather than requiring broad exploratory searches. Accuracy on 16 answer-key questions rose slightly from 0.623 to 0.638, while the graph-assisted system matched or exceeded the baseline on all nine cross-repository dependency questions. Grading combined code-derived answer keys, citation verification, and model judging, although the methodology notes that answer keys cannot fully credit external systems identified by the graph but absent from the checked-out repositories. Results also showed that Haiku did not always include all relevant graph-provided information in its final answers, particularly for telemetry and runtime-call questions, indicating potential gains from improving how the model uses retrieved context.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.