GPT-5.6 Luna's 80% price cut: does cheap Luna change your agent loop?

What it is

GPT-5.6 Luna is the smallest, fastest tier of OpenAI's GPT-5.6 family — the "nano" slot. It's designed for cost-sensitive, high-volume workloads and roughly corresponds to the nano model tier used in earlier GPT-5 families. As of July 30, 2026, OpenAI cut GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%, while adding a premium Fast mode for the flagship.

New rates: $0.20 per million input tokens and $1.20 per million output tokens, a reduction from its previous rates of $1 and $6.

The next-work-session test

You're building an agent loop with a planner-executor pattern. Right now you're paying Sol prices on every tool-selection step, and your evals run costs $40 a pop, so you run them weekly instead of on every PR.

Concrete change: swap Luna in for the router/classifier step and the eval judge. At $0.20/$1.20, a 10K-sample eval that cost ~$40 on the mid-tier now runs for a few dollars. You can put evals in CI. That's the real shift — not "cheaper inference," but evals-per-commit becomes economically defensible.

The catch: you have to actually verify Luna handles the judging task. See Limits.

Pricing

Verified from OpenAI's July 30, 2026 announcement, reported across multiple outlets:

• Input: $0.20 / 1M tokens
• Output: $1.20 / 1M tokens
• Context window: 1,050,000 token context window, maximum output of 128,000 tokens
• Long-context surcharge: Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request
• Prompt caching: Cache writes are billed at 1.25x the uncached input token rate; cached input reads get a 90% discount and the minimum cache life is 30 minutes
• Batch API: Asynchronous batch processing jobs receive up to a 50% discount on both input and output tokens

For long stable system prompts (agent scaffolds, tool schemas), prompt caching stacks with the price cut. The OpenAI model doc is the authoritative reference.

What we'd actually use it for

Narrower than "replace your main model." Honest use cases:

Limits

Luna is a budget tier. Treat it that way.

• Reasoning is weak. GPT-5.6 Luna achieves a score of 27 on the Artificial Analysis Intelligence Index — a composite benchmark that evaluates models across reasoning, knowledge, mathematics, and coding. That's substantially below Sol.
• Agent benchmarks are rough. On ARC-AGI-3 it scores 0% on Public and 0.18% on Semi-Private, and it fails to benefit from extra effort — Luna xhigh is essentially flat. If your agent needs multi-step novel reasoning, Luna is not your executor.
• The long-context surcharge is a trap. Cross 272K tokens and your bill jumps 2x input / 1.5x output on the entire request — not just the overflow.
• Judge-model bias. Using Luna to grade Luna outputs will lie to you. Use it to grade a different model, or spot-check with a stronger judge.
• Benchmarks aren't your workload. Benchmarks should never replace your own evals — the whole reason this price cut matters is that you can now afford to run your own.

Try it if

• You have a router or judge step in your agent loop currently running on an overpriced model.
• You've been putting off building a real eval harness because the token bill scared finance.
• You do overnight bulk processing where Batch API + Luna stacks to ~$0.10/1M input.
• You already have observability (traces, cost-per-request dashboards) to catch regressions from swapping models mid-loop.

Skip it if

• Your agent's core reasoning step is the bottleneck — Luna won't save you, it'll silently make outputs worse.
• You're doing anything ARC-style: novel puzzles, multi-hop planning without strong scaffolding.
• You need a single model to do everything. Luna is a component, not a replacement for Sol.
• Your workloads regularly exceed 272K input tokens — the surcharge erases the discount fast.

The real story isn't "frontier models are cheap now." It's that the cheap tier of a frontier family is now cheap enough to run in your eval loop on every commit. That's the unit-economics shift worth acting on this week.