GPT-5.6 Pricing: What Sol, Terra, and Luna Cost per Token

Key takeaways

• GPT-5.6 ships in three tiers: Sol (flagship), Terra (mid), and Luna (value).
• Reported pricing per 1M tokens: Sol $5 in / $30 out, Terra $2.50 in / $15 out, Luna $1 in / $6 out.
• Output costs 6x input on every tier, so response length, not prompt length, is the real cost driver.
• It is a limited partner preview first, with wider access expected in the following weeks.
• Routing a workload from Sol to Luna can cut token cost by roughly 5x for the same shape of traffic.

What is GPT-5.6, and why three tiers?

GPT-5.6 is OpenAI's newest frontier line, released first as a restricted preview to trusted partners before a broader rollout. The notable part for builders is not the names but the structure: instead of one price, OpenAI is shipping a capability-and-cost ladder. Sol is the flagship, Terra sits in the middle, and Luna is the value option. A tiered family is a pricing signal: OpenAI wants you to match each task to the cheapest tier that still clears your quality bar, rather than paying flagship rates for every call.

How much does each GPT-5.6 tier cost?

Here are the reported list prices, per 1M tokens:

| Tier | Input ($/1M) | Output ($/1M) | Output-to-input ratio | |------|--------------|---------------|-----------------------| | Sol | $5.00 | $30.00 | 6x | | Terra | $2.50 | $15.00 | 6x | | Luna | $1.00 | $6.00 | 6x |

Two things stand out. First, the tiers step down cleanly: Terra is half of Sol, and Luna is roughly a fifth. Second, the 6x output-to-input ratio is identical across all three, which tells you the cost behavior is the same at every tier even though the absolute numbers change.

Why does the output price matter more than the input?

Because output is 6x more expensive than input on every tier, the cheapest lever you have is shorter responses. Say a task reads a long 4,000-token prompt and returns a tight 500-token answer. On Sol that is about $0.02 for input and $0.015 for output, so input dominates. Flip it to a chatty 2,000-token answer and output jumps to $0.06, now four times the input cost. The lesson holds at any tier: trimming verbose completions, capping max tokens, and asking for structured short answers cuts your bill faster than shaving the prompt.

How do the tiers change your unit economics?

Imagine a feature that runs 100,000 calls a month, each with 1,000 input tokens and 1,000 output tokens. That is 100M input and 100M output tokens monthly. Illustratively:

• Sol: (100M x $5) + (100M x $30) = $500 + $3,000 = $3,500/mo
• Terra: $250 + $1,500 = $1,750/mo
• Luna: $100 + $600 = $700/mo

Same workload, same volume, a 5x swing in cost of goods sold purely from tier choice. If that feature is bundled into a $20/mo plan, the tier you pick can be the difference between a healthy margin and serving at a loss. This is exactly the kind of scenario worth modeling before you commit, and you can simulate each tier side by side in Calcaas.

Which GPT-5.6 tier should you actually use?

Default to the lowest tier that passes your evals, then promote only the calls that fail. A common pattern: run Luna or Terra for routine traffic, and route hard or high-value requests to Sol. Because all three share the same 6x output ratio, your prompt and output discipline carries over cleanly when you switch tiers, so you can route by difficulty without re-tuning your cost assumptions.

The takeaway: with GPT-5.6, tier selection is now a first-class pricing decision, so model the per-tier margin impact before access opens rather than after your bill arrives. Want to see it for your own numbers? You can model each tier against your real volume in Calcaas.