OpenAI Just Split GPT-5.6 Into Three Tiers. Here Is Which One Your Business Should Actually Run.

OpenAI previewed GPT-5.6 in early July 2026, and instead of one new model it shipped three. Sol is the flagship. Terra is the balanced middle. Luna is the fast, cheap one. There is also a new maximum reasoning effort for hard problems and an "ultra" mode that spins up subagents for complex, multi-step work.

If you run a business, you do not need the benchmark drama. You need to know which of these to point at which job, and how not to overpay. That is a question we answer for ourselves every week across the platforms we run, so here is the plain-English version.

What OpenAI actually shipped, in plain English

Strip away the naming and it is a tiered lineup, the same shape most labs have moved toward.

• Sol is the flagship, the heavy reasoner meant for the hardest problems.
• Terra is the balanced tier. OpenAI positions it as competitive with the prior GPT-5.5 class at roughly half the cost.
• Luna is the fast, affordable tier for high-volume, lower-complexity work.

On top of the three models, there are two new knobs. A maximum reasoning effort setting lets you tell the model to think harder on a genuinely difficult task. And an ultra mode goes past a single agent, using subagents to break complex work into pieces and run them in parallel. Alongside the models, ChatGPT Business is rolling workspace agents and reasoning-effort controls to admins, with credit-based pricing for those workspace agents beginning around July 6, 2026.

One honest caveat before you build anything on this. GPT-5.6 is a preview rolling out to partners, not a finished, generally available release. Treat it as a direction to plan around, not a production dependency to bet a live workflow on this week. We are writing about it because it is worth understanding early, not because it is ready to carry your traffic today.

Which tier fits which business job

The useful move is to stop thinking about "the best model" and start thinking about "the cheapest model that clears the bar for this specific job." Here is roughly how we would map the three tiers to real work.

| The job | The tier | Why | |---|---|---| | High-volume classification, data extraction, routine drafting | Luna | Simple, repetitive, runs on every request. Cost and speed dominate. | | Everyday reasoning, summaries, customer replies, first-pass analysis | Terra | Needs real reasoning but not the ceiling. The balanced default. | | The rare gnarly plan, a migration, a genuine judgment call | Sol, with higher effort | Hard, low-volume, high-stakes. Worth the flagship and the extra compute. |

The unglamorous rule underneath that table is the one we live by. Most of your production traffic should run on the cheapest tier that passes your evals, not on the biggest model you can afford. The flagship is not the default. It is the escalation.

This is counterintuitive if you think of the top model as "the safe choice." At volume it is the opposite. A flagship you cannot afford to run on every request is not safer, it is just unused, which means most of your actual traffic falls back to something anyway. Better to choose that something on purpose.

Why cheaper usually wins in production

Cost and latency do not add up, they compound. A model that costs twice as much per call and takes twice as long feels fine in a demo and becomes a real problem when you are running it a hundred thousand times a day. A tier that is meaningfully cheaper and still clears your quality bar beats a flagship on every axis that matters once you are at scale, because you can actually afford to run it everywhere.

The new power features cut the same way. Maximum reasoning effort and ultra mode are genuinely useful and genuinely expensive. They are the right tool for the hard five percent of your work, the tasks where a better answer is worth real money and the volume is low. Point them at your whole pipeline and you have built something impressive that you cannot afford to keep on. Scope them to the hard slice and they earn their keep.

None of this is specific to OpenAI. It is the same logic we apply whether the tier is a GPT model, a Claude model, or a Gemini model. More tiers is simply more precision in matching spend to difficulty.

What tiered routing looks like when it is real

This is not theory for us. Across the platforms we run in production, the model choice is never "use the best one." It is "use the one that fits the job, under the constraint that actually matters."

On Smile PreVue, the deciding constraint is not raw capability, it is compliance. The image work runs on Google Vertex AI Gemini under a business associate agreement, because a dental product handling patient data needs HIPAA-grade infrastructure more than it needs a leaderboard score. The tier question there is answered by the constraint before performance ever enters the conversation.

On a messier workload we run, a multi-step attribution and orchestration job, the deciding constraint was reasoning that holds up under genuinely tangled logic, so that piece sits on Anthropic Claude. Different job, different winner, and neither choice was about brand preference. Each was about which model cleared the specific bar that job set.

That is the habit GPT-5.6's three tiers should reinforce, not replace. When a new lineup lands, the question is never "is Sol better than the last flagship." It is "does my extraction job get cheaper if I move it to Luna, does my everyday reasoning hold on Terra, and is there any task in my pipeline that actually needs Sol." You answer that against your own work, one job at a time.

How we would actually adopt this

If a client asked us to bring GPT-5.6 into a real workflow, here is the shape of what we would do.

Start with evals, not vibes. Take a set of your real tasks, the actual tickets or documents or messages your business handles, and run them on Luna and Terra first. Only escalate a task to Sol where the cheaper tier measurably fails. You cannot make this call from marketing benchmarks. You make it from your own work, scored against what "good" means for you.

Route by job, not by brand loyalty. This is the multi-provider point, and it is the one we hold most firmly. The right default for a given task is often a mid or cheap tier from whichever lab happens to win that task, whether that is Claude, Gemini, or GPT. We are not a fan club for one lab. Anthropic, Google, and OpenAI each win different jobs, and pretending one of them wins all of them is how you end up overpaying or underperforming.

Keep an escape hatch. Buy AI a year at a time, not a decade at a time. Design the system so that swapping a tier, or an entire provider, is a config change rather than a rewrite. Models move fast. The preview you are reading about today will not be the newest option for long, and the businesses that stay flexible are the ones who treat the model as a replaceable part, not a foundation poured in concrete. That is the core of how we think about AI integration for business: the plumbing outlasts any single model.

The takeaway for an operator

Strip it all down and the news is good for you. More tiers means you get to match spend to the job instead of paying flagship prices for everything. That is a lever, not a burden.

But the work is still yours. Pick the cheapest tier that passes your evals. Reserve the flagship and the ultra mode for the hard five percent where the answer is worth it. Measure the result against your own tasks, and keep the design loose enough to swap when the next model lands, because it always does.

And it will land soon. Sol, Terra, and Luna are this month's version of a decision you will make again every few months, so the durable win is not picking the right tier today. It is building the muscle, and the wiring, to re-pick it cheaply every time the lineup changes. The businesses that treat model choice as a routine tuning decision, rather than a one-time architecture bet, are the ones who keep getting cheaper and better while everyone else is stuck on whatever they chose a year ago.

If you want help figuring out which tier belongs on which part of your workflow, and wiring it so you are not locked into any one lab, that is exactly the kind of thing we do. Talk to us.