OpenAI's GPT-5.6 Luna dropped to $0.20 per million tokens. Founders, read this before you rip and replace.

What it is

GPT-5.6 Luna is the smallest, fastest model in OpenAI's GPT-5.6 family, pitched for high-volume, latency-sensitive work. On July 30, 2026, OpenAI cut its API price by 80%. When OpenAI introduced the GPT-5.6 series, Luna was priced at $1 per million input tokens and $6 per million output tokens, for a combined total of $7. The new pricing reduces that combined figure to $1.40.

The three sibling models are still around: GPT-5.6 Luna now costs $0.20 input and $1.20 output per million tokens, Terra costs $2 and $12, and Sol is unchanged at $5 and $30.

The next-work-session test

You're the founder. Monday morning, you open your Stripe dashboard and the OpenAI invoice next to it. If your product routes classification, extraction, tagging, first-pass summarization, or the "cheap leg" of an agent chain through GPT-4o-mini or GPT-5.6 Terra, this is the concrete change: swap the model string, rerun your eval set, and your unit economics on that path shrink by roughly 5x versus Luna's old price — or ~10x if you were on Terra.

That's a real GTM lever. Cheaper inference means you can widen the free tier, unlock a per-seat plan that was previously margin-negative, or ship an agent loop that was too expensive to run on every user request. The bet worth making this week: pick one endpoint currently on a more expensive model, run 200 real production prompts through Luna, and diff the outputs.

Pricing

Verified. It's cutting the cost of Luna by 80% to 20 cents per million input tokens and $1.20 per million output tokens. Sol's pricing remains the same. Terra also got a smaller cut — Luna is now 80% cheaper, and Terra is 20% cheaper. Cache writes and reads are billed separately; check the OpenAI model docs for the current rate card and snapshot aliases before you commit budget.

One thing to note if you shop resellers: OpenRouter lists GPT-5.6 Luna at $0.10 per million input tokens, $0.60 per million output tokens. 1,050,000 token context window, maximum output of 128,000 tokens. That's half OpenAI's direct rate, which is worth verifying but not worth architecting around until you know why.

What we'd actually use it for

Honestly? The "cheap leg" of everything.

• Router / classifier at the front of an agent — decide which specialist model handles the request.
• Metadata extraction from user uploads (invoices, contracts, support tickets).
• Draft-then-refine flows where Luna drafts and Sol only edits.
• Bulk backfills — reprocessing your archive of user data for a new feature.
• Chat surfaces where latency matters more than reasoning depth.

Not: your core reasoning path, anything requiring hard math or multi-step planning, or the model your enterprise customers are told they're getting.

Limits

Luna is the small model. Vendor benchmarks are vendor benchmarks — treat them accordingly — but for context, gpt-5.6-luna scores 91.1% on GPQA Diamond, 71.4% on Coding Index, 52.5% on the harder reasoning suites. That's fine for the use cases above. It's not fine for anything where a subtle wrong answer costs you a customer.

The context window is generous — 1.05M token context and 128K token max output — but long-context recall degrades in ways benchmarks don't capture. Run your own needle-in-haystack test on the specific document shapes you care about.

Also: OpenAI cut the price three weeks after launch. That's a signal about competitive pressure, not a promise the price will hold. Don't build a business model that only works at $0.20.

Try it if

• You have a production LLM call that is a bulk, repetitive, or classification-style task and you're currently paying for a bigger model.
• You've been sitting on an agent design that penciled out to "too expensive per user."
• You have a real eval set (even 50 prompts) you can rerun in an afternoon.
• You want to widen a free tier without blowing up gross margin.

Skip it if

• Your product's whole promise is deep reasoning, code generation, or agentic planning — keep Sol on the critical path.
• You don't have evals. Swapping models without them is how quality regressions hit customers before they hit your dashboard.
• Your API bill is under $500/month. The founder-hours to migrate cost more than you'll save.
• You're locked into an enterprise contract that already discounts your current model.

Full price-war context in VentureBeat's writeup and the Forbes take on the AI cost squeeze. The fellou.ai model roundup that surfaced this has the broader landscape.