OpenAI cut prices for its GPT-5.6 family on July 30, 2026 — by 80% for the Luna variant and 20% for Terra. After the change, Luna costs $0.20 per million input tokens and $1.20 per million output tokens, dropping below the cheapest offers from Google and Anthropic. Simon Willison reported the move.
Key takeaways
- GPT-5.6 Luna: $0.20 per million input tokens and $1.20 per million output tokens.
- Luna's cut is 80%; the Terra variant dropped 20%.
- Luna undercuts Gemini 3.1 Flash-Lite (previously $0.25 / $1.50).
- On input, Luna is five times cheaper than Claude Haiku 4.5 ($1 / $5).
- The cut was enabled by GPT-5.6 Sol, which optimized inference on its own.
What changed in the pricing
The biggest move concerns Luna, the cheapest model in the family. The new rate of $0.20 per million input tokens and $1.20 per million output tokens sets it below the previous price leader in this class. Terra, the stronger variant, got 20% cheaper. This is not a temporary promotion but a change to the base API price list.
How it compares with rivals
Until now the cheapest option in the lightweight-model segment was Gemini 3.1 Flash-Lite at $0.25 for input and $1.50 for output. Luna drops below both. The gap versus Anthropic is even sharper: Claude Haiku 4.5 costs $1 for input and $5 for output, making Luna five times cheaper on input tokens. For workloads processing large volumes of text — agents, classification, search — that is a difference that genuinely moves the bill.
| Model | Input | Output |
|---|---|---|
| GPT-5.6 Luna | 0.20 | 1.20 |
| Gemini 3.1 Flash-Lite | 0.25 | 1.50 |
| Claude Haiku 4.5 | 1.00 | 5.00 |
Where the cut came from
OpenAI says it cut costs using its own model GPT-5.6 Sol, which autonomously rewrote compute kernels in the Triton and Gluon languages. The goal was to remove excess memory movement, redundant synchronization and inefficient data layouts that leave GPUs idle. The result is roughly 20% lower end-to-end serving cost. In other words, part of the saving comes not from cheaper hardware but from AI optimizing AI infrastructure.
Simon Willison judged that the move completely changes the landscape in the cheaper-model segment, and migrated his agent demo from Gemini 3.1 Flash-Lite to Luna.
Why it matters
The price of lightweight models determines what is worth building at all. At $0.20 per million input tokens, use cases that made no economic sense with pricier models — mass classification, agents processing long contexts, document-analysis pipelines — suddenly become viable. The fact that the cut is driven by autonomous inference optimization has a second layer: if a model can improve its own serving, the profitability line may shift faster than hardware progress alone would suggest. For rivals it is pressure that is not only about price but about method — Google and Anthropic must respond not with another promotion but with their own efficiency gains. For users it means choosing a cheap model stops being a compromise for the sake of budget.
What's next?
- New GPT-5.6 Luna and Terra prices are live in the OpenAI API as of July 30, 2026.
- The price pressure may force a response from Google and Anthropic in the lightweight-model segment.
Sources
- Simon Willison — Advancing the price-performance frontier with GPT-5.6
