GPT-5.6 Sol vs Terra vs Luna: Which Tier to Actually Use

On June 26, 2026, OpenAI announced the GPT-5.6 family, and for once the naming scheme is doing honest work. Three tiers: Sol, the flagship built for frontier reasoning and long-horizon agentic work. Terra, the balanced everyday model. Luna, the fastest and cheapest of the three. Sun, earth, moon. You are meant to understand at a glance which one is the star of the show, and OpenAI clearly wants you staring at Sol.

I want you to look at Terra instead. But first, one thing you need to know before you plan anything: you cannot use any of this yet. GPT-5.6 is in a limited preview for trusted partners only, available through the OpenAI API and Codex, and there is no general-availability date. Broad release is tied to a White House vetting framework expected around July 7, because models with cutting-edge cyber capabilities now need administration vetting before they go public. So treat this piece as what it is, a planning document, not a buyer’s guide for something you can click into today.

The benchmark crown is real, and it barely matters

Let me give OpenAI its due. On Terminal-Bench 2.1, the agentic coding benchmark that has become the de facto scoreboard for this generation, Sol Ultra scored 91.9 percent and base Sol scored 88.8 percent. That edges out Claude Fable 5 at 88.0 percent and OpenAI’s own GPT-5.5 at 88.0 percent. Sol is, by this measure, the best agentic coding model in the world right now.

By 0.8 points, for the base model. Sit with that number for a second. When Anthropic launched Fable, the story was a genuine step change in long-horizon agentic work. What OpenAI has done here is catch up and nose ahead. That is a real engineering achievement, and it is also the kind of lead that evaporates with the next point release from Anthropic or Google. If you find yourself picking a model vendor because of a sub-one-point gap on a single benchmark, you have already lost the plot. At this altitude, the frontier models are trading blows within the margin of noise, and your actual results will depend far more on your prompts, your scaffolding, and your tolerance for a given model’s quirks than on which one holds the crown this month.

The honest summary is this: Sol proves OpenAI is still in the frontier race. It does not prove you need Sol.

Terra is where the actual news is

Here is the pricing, per million tokens. Sol: $5 input, $30 output. Terra: $2.50 input, $15 output. Luna: $1 input, $6 output.

Terra costs half of what Sol does, and OpenAI says it comes in at roughly half the price of GPT-5.5. That second comparison is the one that should make you sit up. The everyday tier just got about twice as cheap generation over generation, while presumably getting better. That is the actual headline of this launch, buried under the Sol benchmark press.

I keep making this argument and I will keep making it: the mid-tier is where most real work lives. It was true when I wrote about Claude Sonnet 5 becoming the sensible default for agentic work, and it is true here. The flagship exists for a specific kind of user, someone running long autonomous coding sessions, deep multi-step research agents, or workloads where a few percentage points of reliability compound across thousands of unattended steps. If that is you, Sol’s $30 output pricing is defensible, because model cost is a rounding error next to engineer time.

If that is not you, and for most readers of this site it is not, then Terra at $2.50 in and $15 out is almost certainly your answer. Summarization, drafting, customer-facing chat, code review assistance, structured extraction, the unglamorous 90 percent of what companies actually deploy. None of it needs the last 3 points of Terminal-Bench.

And Luna, at $1 in and $6 out, is the tier I suspect will quietly move the most tokens. Classification, routing, autocomplete-style tasks, anything high-volume and latency-sensitive. The pattern that wins in production is boring and effective: Luna handles the flood, Terra handles the substance, and Sol gets called only when a task genuinely demands it. If your architecture cannot route between tiers like that, fixing that is worth more than any model upgrade.

What to actually do before it ships

Since none of us outside the preview can touch these models yet, here is how I would spend the wait.

• Instrument your current usage. You cannot decide between tiers if you do not know your input-to-output token ratio and which tasks eat your budget. Most teams I talk to are guessing.
• Default your plan to Terra. Make Sol earn its way into your stack with evidence from your own evals, not from OpenAI’s launch charts.
• Do not migrate anything on day one. The vetting framework around July 7 is new territory, and I would not bet a production system on a launch date that depends on a government process hitting its schedule.

My honest read: GPT-5.6 is a strong, slightly unexciting launch dressed up as a dramatic one. The Sol versus Fable rivalry will dominate the headlines because horse races are fun to write about. But the decision that will actually touch your budget is Terra versus whatever mid-tier you run today, and on the numbers OpenAI has shown, that fight just got genuinely interesting. Plan for that one, and let the flagships fight over their 0.8 points without you.