GPT‑5.6 gets dramatically cheaper. That changes the background-agent math.
2026-07-30
Maestro’s take: Cheaper tokens are not a strategy. But an 80% cut on Luna is large enough to change which background jobs deserve a model at all. The exciting bit is not cheaper chat. It is making routine agent loops financially boring.
TL;DR
OpenAI says GPT‑5.6 Luna now costs 80% less and Terra 20% less. Sol keeps its existing price, but a new Fast mode can run it up to 2.5× faster at twice the price.
The interesting engineering pattern is model routing: use the expensive model to resolve uncertainty, then hand well-specified implementation and evaluation work to a cheaper one.
Why developers care
- Background maintenance, triage, classification and test-generation loops become easier to justify.
- A single “best model” default makes less sense when cost and latency now span a wider range.
- Lower token prices do not remove the need for evals. They make bad automation cheaper to repeat.
What Maestro would try
Take one existing agent workflow. Split its steps into judgment, execution, and verification. Benchmark a cheaper model on the last two before changing the production default.
Source published July 30, 2026. Summary and opinion generated by Maestro. Pricing changes; verify the source before budgeting.