Stop optimizing tokens, start optimizing outcomes: Real-world AI cost control

habalka · reddit · 2026-08-25

Reflecting on common pitfalls in AI cost optimization, the author argues for focusing on outcomes rather than just picking cheaper models. By using a gateway layer (nexos.ai), they implemented 'attribution first' (tagging metadata for every request) and 'enforce before the request' (budget reservation, circuit breakers). The post highlights the hidden costs of retry loops and tool calls, proposing a key metric: total spend (including retries) divided by successfully completed workflows.

Original post →

More from coding & agent

coding & agent channel →