Blog · 2026-07-10

GPT-5.6 Sol vs Terra vs Luna: Which Tier Should You Actually Use?

OpenAI's GPT-5.6 family ships in three tiers — Sol, Terra and Luna — and they are the same model line at three price points, not three different architectures. All three take text and image input, return text, carry a 1,050,000-token context window with up to 128,000 output tokens, support reasoning tokens, and share a February 2026 knowledge cutoff. Choosing between them is almost entirely a cost decision.

On apimodels.app the three sit at a clean 5 : 2.5 : 1 ratio. Sol costs $1.103 per million input tokens and $6.618 per million output. Terra is exactly half of that, $0.551 and $3.309. Luna is a fifth of Sol, $0.221 and $1.324. Cached input tokens are billed at a tenth of the input rate on every tier — $0.110, $0.055 and $0.022 respectively. For comparison, OpenAI's own list prices for the same three models are $5/$30, $2.5/$15 and $1/$6, so each tier lands at roughly a fifth of official.

A practical rule: start on Terra. It is the tier most workloads should default to — everyday chat, code generation, agentic tool calling — because doubling your spend to reach Sol buys reasoning depth you often will not use. Move up to Sol when a task genuinely fails on Terra: long multi-step planning, hard debugging, research synthesis. Move down to Luna when the job is bulk and mechanical: classification, extraction, rewriting, first-line support. At $0.221 per million input tokens, Luna makes batch pipelines that would be uneconomical on a frontier model suddenly cheap.

Reasoning depth is not a body parameter on GPT-5.6 — it is a suffix on the model name. gpt-5.6-sol-low, -medium, -high, -xhigh, -max and -ultra are six distinct model ids, and crucially they all cost the same per token. Higher effort simply emits more reasoning tokens, which are billed as output. So -ultra is not a more expensive model; it is a model that tends to produce a longer, costlier response. That also means -max and -ultra are not legal values for the reasoning.effort field used by earlier GPT-5 models: you switch depth by changing the model string, not the request body.

Three behaviours are worth knowing before you build on any of the three. First, upstream injects a system preamble of roughly 1,450 tokens into every request, so usage.input_tokens never drops below that floor even for a one-word prompt — budget for it when you price short, high-frequency calls. Second, parallel_tool_calls: false is ignored: ask about the weather in Paris and Tokyo and you will get two tool_calls back regardless, so write your handler to iterate rather than assume one. Third, once a single request exceeds 272,000 input tokens the entire request is rebilled at 2x input and 1.5x output — the step is on the request, not on the tokens past the threshold.

All three answer on both OpenAI-compatible endpoints, /v1/chat/completions and /v1/responses, with function calling, streaming and multi-turn tool results verified on each. Point base_url at https://apimodels.app/api/v1, keep your existing OpenAI SDK, and switch tiers by editing one string. Two aliases exist for convenience: gpt-5.6 routes to Sol, and gpt-5.6-luna routes to gpt-5.6-luna-max, which is currently the only Luna depth the upstream serves — same price, so nothing changes on your bill.

FAQ