OpenAI's GPT-5.6 Comes in Three Tiers. Picking the Right One Is the New Skill.

On 9 July, OpenAI launched the GPT-5.6 family in three variants: Sol, the workhorse; Terra, a middle option; and Luna, the budget model. Pricing per million tokens runs about $5 in and $30 out for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. OpenAI calls Sol its strongest cybersecurity model yet and reports a score of 80 on the Artificial Analysis Coding Agent Index while using fewer output tokens than the last generation. All three are available through ChatGPT, Codex, and the API.

Why the tiering matters more than the benchmark

A single frontier model forces one price on every task you run. A tiered family lets you spend Sol money only where a task truly needs it and fall back to Luna everywhere else. Across Sol to Luna that is roughly a fivefold gap on both input and output, which for anyone running models at volume is the difference between a healthy cost per task and a bill that quietly runs away.

So the skill shifts. It is no longer use the best model. It is use the cheapest model that still clears your bar, and be able to prove it clears it. That proof is the hard part, and it is the same discipline a real LLMOps stack exists to give you: a way to score each task against a standard instead of trusting a hunch. A benchmark score is not your workload. Sol scoring 80 on a coding index tells you nothing about whether Luna is good enough for your extraction step or your router.

The catch nobody prices in

Build on a model and you inherit its lifecycle. OpenAI is retiring o3 from ChatGPT on 26 August and the DALL-E GPT on 30 August. Every model you depend on has a sunset, and when it arrives you re-test and re-qualify on its replacement. If you run agents where cost compounds, that churn is a standing tax, and it shows up plainly once you look at what an AI agent actually costs to run.

What to do with it: do not standardise on one model. Default to the cheapest tier, and escalate to Sol only when your own check says the cheap one failed. Write the check first. If you cannot say what good enough means for a task, you cannot safely pick a tier, and you will overpay for the top model out of fear.