The original title is "I Built a Task Orchestrator, Then Deleted Its Best Number"

Short summary

ORA is an open-source Go binary that decomposes tasks into subtasks and routes each to the cheapest model tier that can handle it, working with Claude Code, Codex, Cursor, and others. The author removed a misleading '68% cheaper' savings claim after realizing it was a counterfactual story, not a measurement — replacing it with actual run cost. The post is a candid lesson on why most 'X% cheaper' AI tooling benchmarks are assumptions dressed as data.

• •ORA routes subtasks to cheapest capable model tier, reconciles results
• •Author deleted fake 68% savings metric, replaced with actual run cost
• •Argues most AI tooling savings claims are counterfactual stories, not measurements

Generated with AI, which can make mistakes.