If you read one line: OpenAI cut Luna 80% and Terra 20% on July 30. On the sticker Luna is now the cheapest agent backend alive — but you pay per token and ship per completed task, and cheaper tiers spend more tokens per task. Route on cost-per-completed-task, and the ranking barely moves.

What actually changed#

On July 30, 2026, OpenAI repriced two of the three GPT-5.6 tiers (OpenAI, Finout):

Nothing about the models changed — same weights, same context, same three-tier menu OpenAI shipped on July 9. Only the number on the invoice moved. And it moved into territory that reads, at a glance, like a market event: Luna at $0.20 / $1.20 now sits roughly 7× under Gemini 3.6 Flash's $1.50 / $7.50 on input and 6× under on output, and about 15× under Kimi K3's $3 / $15 (kie.ai, eesel). Ten days ago Flash was the cheap workhorse everyone routed to. On the pricing page, that's over.

The number on the page is not the number on the bill#

Here's the trap, and it's the same one we've flagged every time a price war reprices the cheap tier: you are billed per token, but your product ships per completed task. Those are different units, and the discount lives in the wrong one.

The bill for a task is:

cost = (tokens spent to finish the task) × (price per token)

A price cut moves the second term. But smaller, cheaper tiers routinely lose on the first term — they spend more reasoning tokens, more tool-call rounds, and more retries to reach the same finished result. We measured exactly this shape when Grok 4.5 won the bill while losing the benchmark: the decisive variable was the token count, not the token price. Run it the other way and the same physics bites you — a tier that's 6× cheaper per token but takes 2-3× the tokens (and the occasional full retry) to complete a long-horizon agent task can hand you a bill that's a rounding error away from where you started.

The unit that routes correctly is cost per completed task. Everything you can read off a pricing page is the wrong unit.

This is why the resolution, not the token, is the unit you should be optimizing. Sol's price didn't move on July 30, and that's the tell: OpenAI didn't need to discount the tier that posts a 53.6 on Agents' Last Exam (Vellum). Frontier completion rate is its own pricing power. The discount went to the tiers that have to earn their place on your task, not their place on a leaderboard.

The routing map that survives the cut#

Don't migrate on the sticker. Migrate on your harness.

  1. Pin your evaluation suite first. Before you touch a route, make sure you have a fixed set of your real agent tasks with a pass/fail bar. If you don't, that's the actual project this week — not the model swap. (If you run Langfuse or a trace-and-eval setup, you already have the substrate.)
  2. Run Luna against it, cold. Measure completion rate, tokens-per-completed-task, and tool-call rounds — not vibes on three prompts. Compute the real cost per passed task, retries included.
  3. Promote only if completion holds. If Luna finishes your suite within tolerance, the 80% cut is a genuine, large win — take it, and cross-check that its cheaper fast mode doesn't quietly drop quality on your hardest cases.
  4. If it doesn't, look at Terra, not Flash. The quieter story is Terra's move to $2 / $12 — a 20% cut on the balanced tier most founders actually run production agents on. That's the mid-tier decision that shifted without a headline.
  5. Keep Sol for the hard horizon. Unchanged price, unchanged job: the long-running tasks where a re-run costs more than the frontier premium.

What it means for you: the July 30 cut is real and, for the right workload, large — but it's an invitation to measure, not a reason to swap routes on faith. The founders who win the next month aren't the ones who read the cheapest number off a pricing page. They're the ones who already knew their cost per completed task before OpenAI changed the sticker, and could tell within a day whether the discount survived contact with their own workload.