GPT-5.6 Pricing Explained: Sol vs Terra vs Luna Cost Breakdown

Key takeaways

• Three tiers, one generation: Sol ($5 in / $30 out), Terra ($2.50 / $15), Luna ($1 / $6) per 1M tokens.
• Terra is positioned as competitive with GPT-5.5 at about 2x cheaper, resetting the default for everyday work.
• The Sol-to-Luna spread is roughly 5x, so tier routing affects margin more than family choice.
• New caching rules: cache writes cost 1.25x input, cache reads keep a 90% discount, with a 30-minute minimum cache life.
• A cached prefix pays off after about two reads, making caching a real lever for repetitive prompts.

What does GPT-5.6 cost?

OpenAI previewed the GPT-5.6 series with three durable tiers. Per 1M tokens:

• Sol (flagship): $5 input, $30 output
• Terra (balanced): $2.50 input, $15 output
• Luna (fast, low cost): $1 input, $6 output

Output is six times the input price across all three tiers, which matters because most real workloads turn output-heavy once you add reasoning and tool calls.

Which tier should you use?

Terra is the headline. OpenAI positions it as competitive with the prior GPT-5.5 while costing about half. For a lot of everyday product work, that makes Terra the new default, with Luna for high-volume or latency-sensitive paths and Sol reserved for the hardest reasoning.

The bigger point is structural: the spread from Luna to Sol is roughly 5x on both input and output. A product that routes each task to the cheapest tier that still passes quality will beat a product that hardcodes one model, regardless of which family it chose. Tier routing is now a first-class pricing decision.

How does the new prompt caching math change costs?

GPT-5.6 introduces more predictable caching: explicit cache breakpoints, a 30-minute minimum cache life, cache writes billed at 1.25x the uncached input rate, and cache reads keeping the 90% cached-input discount.

Here is the break-even, framed illustratively. Say a fixed prompt prefix would otherwise cost 1 unit of input each time you send it. Without caching, N sends cost N units. With caching, you pay 1.25 once to write, then 0.1 per read. Caching wins when:

1.25 + 0.1N <= N

That solves to N of about 1.4. In plain terms, if a cached prefix is reused roughly twice or more inside the 30-minute window, caching is already cheaper. For chat histories, system prompts, and RAG context that repeat constantly, that threshold is trivial to clear.

What about throughput?

OpenAI also flagged GPT-5.6 Sol on Cerebras at up to 750 tokens per second in July. Faster output does not change the per-token price, but it cuts user-perceived latency and lets more work fit into time-boxed agent runs, which indirectly affects how many tokens a task consumes.

How do you turn this into a margin decision?

Pricing tables do not tell you your cost. Your token mix does. To compare GPT-5.6 tiers against your current provider:

1. 1Pull your average input and output tokens per task.
2. 2Apply each tier's input and output price for a blended cost per task.
3. 3Layer in caching for any repeated prefix using the break-even above.
4. 4Multiply by tasks per active user to get cost per user, then compare to your price.

Takeaway: GPT-5.6 makes tiering and caching the main cost levers, not the brand on the API. Model your token mix across Sol, Terra, and Luna in Calcaas before you commit a default.