OpenAI Slashes Luna Pricing 80%, but Agent Costs Barely Budge
An 80% model price cut creates an illusion of proportional savings that collapses the moment tool calls, retries, and human review enter the ledger. Teams that migrate on token price alone will over-provision Luna, under-budget for Web Search and long-context step jumps, and discover the real cost per successful task barely moved.
OpenAI's July 2026 pricing update reduces Luna Standard input, cache read, cache write, and output rates to exactly 20% of their previous levels. For a homogeneous token bill, the math is clean: costs drop 80%. But agent workloads are never just tokens. Tool calls, loop retries, long-context step jumps at 272K tokens, cache mismatches, and human review all add independent costs that the model price cut leaves untouched. A task that once cost $0.20 might only fall to $0.12 — a 40% reduction, not 80%.
The real metric to track is cost per success, not cost per token. Web Search fixed fees now exceed the model token cost for a typical request, making tools the new cost center. The 272K context threshold triggers a discrete rate jump to 2× input and 1.5× output pricing, so average context figures hide dangerous tail behavior. Cache architecture matters too: writing a prefix that never gets read is pure waste at 1.25× the normal input rate.
A practical migration path runs 100–500 real tasks through a Luna-first routing tier with conditional escalation to stronger models. Low-uncertainty, auto-verifiable tasks go straight to Luna. High-stakes, hard-to-verify tasks stay on Terra or Sol with human review. The 30-day plan starts with baseline exports, moves through small-traffic A/B tests, and ends with hardened upgrade and rollback thresholds.
Token price cuts shift the cost bottleneck from the model to the tooling layer. When a single Web Search call costs more than the model inference for a typical request, agent architecture must optimize tool call frequency, not token count.
The 272K threshold is a billing cliff, not a gradual slope. A request at 273K tokens pays double input rates on the entire request, so teams that only monitor average context length will miss catastrophic tail costs.
Cache write pricing at 1.25× normal input inverts the usual intuition that caching is always cheaper. Writing a prefix that never gets a cache hit is strictly more expensive than sending it as uncached input.
Cost per success exposes a fundamental tension: cheaper models may require more retries, more tool calls, and more human review, erasing the per-token savings. The metric forces teams to optimize the whole pipeline, not just the model choice.
Benchmark scores are useless for migration decisions because they exclude tool loops, long-context step jumps, and error costs — the very factors that dominate real agent task economics.