GPT-5.6 Is Here
OpenAI's current model guidance documents GPT-5.6 as a three-tier family: Sol for frontier capability, Terra for a balance of intelligence and cost, and Luna for efficient high-volume work. This article uses the current model documentation and does not rely on an unverified rollout story.
The New Naming: Sol, Terra, Luna
With GPT-5.6, the number is the generation and the name is the capability tier:
- Sol — the frontier tier for complex professional work. The
gpt-5.6alias routes togpt-5.6-sol. - Terra — the balanced tier for strong performance at a lower price.
- Luna — the cost-sensitive tier for high-volume workloads.
Pricing (per 1M tokens)
| Tier | Input | Output | Positioning |
|---|---|---|---|
| Sol | $5.00 | $30.00 | Frontier capability |
| Terra | $2.00 | $12.00 | Balance of intelligence and cost |
| Luna | $0.20 | $1.20 | Cost-sensitive, high-volume work |
OpenAI's model comparison lists cached-input prices of $0.50, $0.20, and $0.02 per million tokens for Sol, Terra, and Luna respectively. These are the short-context standard rates; OpenAI documents a higher long-context tier above 272K input tokens. Recheck those conditions before forecasting a large-context workload.
What's Actually New
- New tier names. The family separates frontier, balanced, and high-volume roles without forcing every request through one model.
- Long context. OpenAI currently lists a 1.05M-token context window and 128K maximum output for all three tiers.
- Reasoning control. The documented reasoning levels run from
nonethroughmax; test the existing setting and one lower level during migration. - Tool workflows. Current model pages list function calling, structured output, streaming, and image input. Verify the exact endpoint and tool support used by your application.
How to Call GPT-5.6
OpenAI recommends the Responses API for reasoning, tool-calling, and multi-turn workflows. If LinkModel lists the same model ID, confirm its compatible endpoint and fields in the LinkModel documentation before using this Chat Completions-shaped example:
curl -X POST https://api.linkmodel.ai/v1/chat/completions \
-H "Authorization: Bearer $LINKMODEL_API_KEY" -H "Content-Type: application/json" \
-d '{ "model": "gpt-5.6-terra", "messages": [{"role":"user","content":"Refactor this module and add tests: ..."}] }'Confirm the exact model strings in the docs. New to keys? See how to get an OpenAI API key.
Which Tier Should You Use?
- Hardest coding / agents → Sol (but $30 output — reserve for the hard 5%).
- Everyday production → Terra ($2/$12 at the short-context standard rate).
- High-volume / cheap → Luna ($0.20/$1.20 at the short-context standard rate).
Route the bulk to Luna/Terra and escalate to Sol only when needed — the tiered pattern in how to reduce AI API costs. Compare against rivals in best coding LLM API and Claude vs GPT; full rates in GPT API pricing.
What to Verify Before Treating This as a Release Guide
This page is intentionally dated. Before shipping code, confirm four items in OpenAI's current model documentation: whether each tier is available to your project, the exact API model ID, the current token rates, and which reasoning or tool features each tier accepts. A name appearing in a third-party catalog is not proof of direct API availability.
A Safer Tier-Selection Test
Run the same 30–50 production tasks through each available tier. Score task success before latency or token price.
| Signal | Why it matters |
|---|---|
| Pass rate | exposes false savings from cheap failed attempts |
| Median and p95 latency | catches long-tail delays hidden by averages |
| Input/output tokens | shows whether a model is concise or burns budget |
| Tool-call errors | matters for agents more than chat benchmarks |
| Human repair time | converts quality gaps into operational cost |
Calculate cost per accepted task = total model spend ÷ accepted tasks. Escalation works when a lower tier handles routine work and uncertain or failed cases move upward. It fails when the router sends every ambiguous request to the expensive tier.
Migration Checklist
- Put the model name in configuration, not application logic.
- Capture response and tool-call schema differences in tests.
- Set per-request token and step budgets.
- Keep a rollback model during preview periods.
- Recheck safety, retention, and regional terms for the API surface you use.
Do not migrate solely from a vendor benchmark. A benchmark can justify a test; it cannot substitute for one.
Bottom Line
GPT-5.6 splits into Sol (frontier coding), Terra (value default) and Luna (cheap, punchy). If you build coding agents, Sol's efficiency is the story; for everyday work, Terra is the sweet spot. Test it against Claude and Gemini on your own prompts before committing.
Start free with a $1 credit and benchmark GPT-5.6.