GPT-5.6 Luna's 80% price cut: what actually changes in your agent loop

What it is

GPT-5.6 Luna is the small, fast tier of OpenAI's GPT-5.6 family, sitting under Terra and Sol. It is suited for high-volume, latency-sensitive tasks such as chat, classification, and lightweight agent work. On July 30, OpenAI dropped its API price — the headline reason it's worth a fresh look for anyone shipping agents.

The next-work-session test

You've got a research agent that plans, calls three tools, drafts, then self-critiques. On Sol, one full loop over a 40k-token working set was too expensive to run in CI, so your evals sample 20 traces a night.

At Luna's new price, the same loop drops roughly an order of magnitude. Practical change: you can move judge and plan calls onto Luna, keep Sol only for the final synthesis step, and run the full eval suite on every PR instead of nightly. That's the concrete win — not "cheaper AI," but "evals move left."

Pricing

Verified against OpenAI's own announcement: "Starting July 30, API pricing is $2 per million input tokens and $12 per million output tokens for Terra, and $0.20 per million input tokens and $1.20 per million output tokens for Luna," per OpenAI's price-performance post. InfoWorld reported the cuts as up to 80% across Luna and Terra.

One caveat worth flagging: third-party aggregators list different numbers. OpenRouter shows $0.10 per million input tokens, $0.60 per million output tokens for its Luna route, which is a reseller price and not what you'll pay hitting OpenAI directly. Check whose endpoint you're actually billing against.

What we'd actually use it for

• The cheap half of a two-model agent loop. Route planning, tool-argument formatting, retrieval query rewriting, and pass/fail judging to Luna. Reserve Sol or Terra for the reasoning-heavy step.
• Eval harnesses. LLM-as-judge over thousands of traces gets tractable. Pair with a stronger model on a stratified sample to keep judges honest.
• Classification and extraction at volume. Ticket triage, log tagging, structured-field extraction from documents.
• Long-context skims. The big context window helps here, with the caveat below.

Limits

The context window is real but shallow at the tail. One review harness noted Luna started summarizing instead of citing at around 300k tokens, which is fair behavior for a $1 model but worth knowing before you architect around it. If your agent depends on faithful long-document quotation, don't just swap Sol → Luna and assume parity.

On raw capability, Luna is not intended to win flagship benchmark charts. Its value is that it retains the 1.05 million-token context window and core GPT-5.6 features at one twenty-fifth of Sol's input and output price. Translation: for genuinely hard multi-step reasoning, you'll still want Sol in the loop somewhere.

Also — the vendor's own pricing chart is the ground truth. Aggregator dashboards drift, and a "blended" cache-hit rate (Artificial Analysis quotes $0.17 per 1M blended) only applies if your traffic actually hits cache. Agent loops with high prompt diversity often don't.

Try it if

• You run an agent loop where >60% of tokens go to planning, judging, or formatting — the "cheap half" is exactly Luna's shape.
• Your eval pipeline is throttled by cost, not signal quality.
• You do high-volume classification or extraction and Sol was overkill.
• You want a cheap draft model to pair with a stronger reviewer.

Skip it if

• Your workload is a single-shot hard-reasoning call. Just use Sol; the price delta on one call doesn't matter.
• You need faithful long-context citation past ~300k tokens.
• You're already on a competing frontier-cheap tier (Claude Haiku, Gemini Flash) and your evals show parity — the switching cost isn't free, and per-token math alone rarely justifies a re-plumb.
• You haven't measured your current token mix. "80% off" on a workload that's 90% output tokens is a different discount than on one that's 90% input.

The real move this week isn't switching models. It's re-running your token accounting with the new numbers, then asking which step in your agent loop deserves a demotion.