GPT-5.6 explained: what Sol, Terra, and Luna mean for your business

On June 26, 2026, OpenAI previewed its next model family, GPT-5.6, in three tiers named Sol, Terra, and Luna (OpenAI). At preview it went to roughly 20 organizations, shared with the US government first, and almost nobody could call it (Axios).

Thirteen days later, on July 9, it was generally available across ChatGPT, Codex, and the API. In the six weeks after that, the price fell twice.

That sequence is the actual lesson of this release, and it is worth more than the benchmark numbers. Here is what the tiers are, what they cost now, and what a business should do about it.

What OpenAI shipped

GPT-5.6 is not one model. It is three, split by job and price:

• Sol is the flagship, tuned for the hardest reasoning and agent work.
• Terra is the balanced everyday model.
• Luna is the fastest and cheapest, for high-volume, latency-sensitive work.

Current list prices per million tokens, against GPT-5.5 for comparison:

GPT-5.5 remains $5 input and $30 output (OpenAI API pricing).

Read the top row twice. The new flagship costs less than the model it replaces. That is not how flagship launches usually go, and it changes the arithmetic on work you priced out earlier this year.

The prices moved twice after launch

The preview numbers were Sol $5 in and $30 out, Terra $2.50 and $15, Luna $1 and $6. None of the three survived the summer:

• July 30: Terra fell 20% and Luna fell 80%.
• August 21: Sol fell to $4 and $20, a promotional rate OpenAI has said will hold at least until November 21, 2026.

So a business case built on the June preview numbers was wrong by August, and wrong in the direction of overestimating cost. Anything you read about GPT-5.6 pricing that predates July 30 understates how cheap the lower tiers now are.

Note the word promotional on the Sol row. Sol is the one tier whose current price has a published end date. If you are modelling a twelve-month bill on Sol, model the $5 and $30 rate returning, and treat the discount as upside rather than as your baseline.

The names, in plain terms

The tier names are marketing, but the pattern behind them is now standard across the industry. Every major lab ships a flagship, a mid-tier, and a cheap fast option. Anthropic has Opus, Sonnet, and Haiku. Google has Gemini Pro and Flash. OpenAI has Sol, Terra, and Luna.

Translate the names to jobs and the decision gets easier: Sol for the hardest reasoning and agent tasks, Terra for most day-to-day work, Luna for high-volume, cost-sensitive calls.

This matters because most teams overpay by sending every request to the flagship. In practice a large share of production traffic is routine and runs fine on the cheaper tier. Choosing the right tier per task, rather than defaulting to the top model, is the single clearest way to control cost. For the full vendor-by-vendor picture, see our guide to choosing the right LLM for your business.

What this means for your AI budget

With Luna at $0.20 and $1.20, the gap between tiers is now wide enough that routing is worth engineering properly rather than treating as a nice-to-have.

Here is a labelled hypothetical to make that concrete. It is illustrative, not a client result. Consider a support team handling 10 million tokens of routine ticket classification a month, plus a smaller slice of hard escalations. Sending all of it to the flagship is the expensive default. Routing the routine classification to Luna and reserving Sol for the hard escalations changes the bill by more than an order of magnitude on the routine portion, because the tiers are now 20 times apart on input rather than five.

The exact saving depends on your traffic mix, which is why the routing design, not the headline model, is where the money is. Our AI agent cost and pricing guide walks through how to run that math on real volume.

What to do right now

Four steps, in order.

Re-run any business case built before July 30. The lower tiers are substantially cheaper than they were at preview. Projects that did not clear the bar on the June numbers deserve a fresh calculation rather than a second opinion.

Route by task, not by habit. Measure what share of your traffic is genuinely hard. For most teams it is a minority, and the rest can move down a tier without a quality change your users would notice. Test that claim on your own prompts before you believe it.

Keep the model swappable. If your system can change models with a config change rather than a rewrite, the next reprice is an afternoon. If your prompts and pipelines are hard-wired to one model, every release is a project. This release is the argument: three price changes in eight weeks, and the teams that captured them were the ones who did not have to rebuild anything.

Model Sol at its standard price. The current Sol rate is promotional through at least November 21, 2026. Budget on $5 and $30 returning, and bank the difference until OpenAI says otherwise.

The bottom line

GPT-5.6 went from a preview almost nobody could call to generally available in thirteen days, then got cheaper twice in six weeks. The flagship now costs less than the model it replaced, and the cheapest tier costs a twentieth of it.

The durable lesson is not which tier wins. It is that model prices are moving faster than most teams' architectures can follow. The teams that win the next pricing cycle are the ones whose systems can adopt a cheaper model the week it ships. For the open-weight side of that comparison, our guide on open source versus proprietary models covers the trade-offs.