GPT-5.6 Is Three Models Now: Route Your Prompts Across Sol, Terra, and Luna (July 2026)

On July 9 OpenAI shipped GPT-5.6, and the news is not the benchmark, it is the shape. It is not one model. It is three durable tiers at three prices, Sol, Terra, and Luna, plus two new reasoning dials, max and ultra. The reflex is to swap your model string to gpt-5.6 and move on. That is the expensive mistake. The tier is now a per-task routing decision that persists across versions, and it only pays off if your prompts are versioned and labeled by the tier they were tuned for. Here is the routing framework, when the new effort dials earn their cost, and what it changes for your library.

What shipped on July 9 (the shape, not the score)

The naming is the tell. In OpenAI's new system the number is the generation and the name is the durable tier: Sol is the flagship, Terra is the balanced everyday tier, Luna is the fast, low-cost tier, and each advances on its own schedule. When 5.7 lands, Sol/Terra/Luna stay; only the number moves.

The prices are the reason it matters. Sol is $5 in /$30 out, Terra is $2.50 /$15, and Luna is $1 /$6 per million tokens, all with a roughly 1M-token context window. Two new reasoning-effort settings sit on top: max gives a single agent the most time to reason, and ultra spins up subagents to push past what one agent can do. Caching got more predictable too, with cache writes billed at 1.25x the uncached input rate and cache reads keeping the 90% discount.

The benchmarks are strong (Sol leads Terminal-Bench 2.1 at 88.8%, and sets an Agents' Last Exam high of 53.6), but they are not the point for a prompt library. The point is that OpenAI just handed you a 3-by-N grid of prices and effort levels where you used to pick one box.

"Just point it at gpt-5.6" is the expensive reflex

One price does not fit every task. Run a high-volume classification loop on Sol and you pay roughly five times Luna's rate for output no one can tell apart. Run a multi-step refactor on Luna and you save money right up until it quietly gets the third step wrong. The tier is a per-task decision, not a project default.

We have made this argument before across vendors, when a cheap agentic model crossed the usability line: default down to the cheap tier for the volume, escalate up to the frontier for the hard 10%. GPT-5.6 moves that same decision inside one family, which makes it both easier (one API, one tokenizer) and stickier: because the tiers are durable, the routing you set up this week survives 5.7 and 5.8. This is the model-agnostic discipline from choosing the right model in 2026, except now the "models" share a namespace.

The routing framework, tier by task

The line is not "which is best," it is "which is enough." Match the tier to the task's stakes and specification, not to your ambition. When we re-routed our own library, the boring high-volume work, the classification and extraction loops, dropped to Luna and nothing downstream noticed except the bill; the judgment calls stayed on Terra and Sol.

• Luna ($1 /$6): the volume floor. Well-specified, low-stakes, high-frequency work: classification, extraction, formatting, routing, first drafts you will edit. If the task has a clear right answer and a tight output contract, Luna almost certainly clears it.
• Terra ($2.50 /$15): the everyday default. Drafting, summarization, most agent steps, code you will review before it ships. This is where the bulk of interactive work lives.
• Sol ($5 /$30): the hard tail. Multi-step reasoning, long agentic runs, high-stakes correctness, the 10% of calls that actually decide an outcome. Pay for Sol where being wrong is expensive, not everywhere.

The routing matrix. Route by the task's stakes and how tightly it is specified, not by which tier scores highest. Most of a library lands on Luna and Terra; Sol is for the tail.

Do not take our split as gospel. Let your own tests draw the line: run the same prompt on Luna and Terra, and only promote it to a higher tier when the cheaper one measurably fails the task, not when it feels safer.

The second axis: max and ultra

Effort is orthogonal to tier. You can run Terra at standard effort or Sol at ultra, and the cost and latency change accordingly. max buys a single agent more reasoning time; ultra decomposes the work across subagents.

Spend it where the task is genuinely multi-step and decomposable and the answer matters: a migration plan, a gnarly refactor, a long agentic run with branching. Do not spend it on the other 90%, where it just adds latency and tokens. This axis rhymes with the reasoning-effort shift GPT-5.4 introduced last winter: the dial is not a quality slider you crank for comfort.

The trap is using max or ultra to paper over a vague prompt. Effort makes a well-specified prompt think harder; it makes an underspecified one improvise harder. If a prompt is not producing what you want, score its structure first and fix the missing context or format before you pay for deeper reasoning. Effort is the last dial you turn, not the first. We learned that the slow way, cranking ultra on a summarizer that only ever needed an output format and paying for subagents to over-think a one-line answer.

The effort dial. Standard covers most calls; max and ultra are for genuinely multi-step, high-stakes work. Fix the prompt before you buy more reasoning, or you are paying to improvise harder.

What it changes for your prompts

Here is the part that outlives this release. A prompt is now a (tier, effort) artifact, not a free-floating string. The same prompt behaves and costs differently on Sol versus Luna, and a prompt written for one tier is not automatically right for another.

A prompt padded with scaffolding to compensate for a weaker model wastes money on Sol, where the reasoning is already there, and may still limp along on Luna. A lean prompt that quietly assumed Sol-level reasoning can underdeliver the moment you route it to Luna to save money. Neither failure is visible until you measure it. The prompts that broke when we first routed down were exactly the ones we had over-engineered months earlier to prop up a weaker model; the lean, well-scored ones ported without a blink.

So the discipline is concrete: version each prompt with the tier it was tuned for, score it, and re-check the score when you re-route it to a cheaper tier. Portability is not "runs on everything," it is "runs acceptably on the tiers you would actually route it to." This is also why Terra matters beyond OpenAI: at $2.50 /$15 it undercuts Claude Sonnet 5's introductory pricing, so the cross-vendor routing question is live again, and a portable, scored library is what lets you answer it in an afternoon instead of a sprint.

One prompt, three tiers. Version and score the prompt for the tier it targets, because the same text delivers different quality and cost on each. Re-check the score before you route it down to save money.

Do this week

The migration is mostly a routing and labeling job, not a rewrite.

1. Route the volume down. Move well-specified, low-stakes calls to Luna or Terra. This is where the bill actually lives.
2. Reserve Sol plus max/ultra for the hard tail. The 10% of calls where being wrong is expensive.
3. Re-measure tokens and caching. The new cache-write surcharge (1.25x) and the 90% read discount change the math on cache-heavy workloads; confirm before you assume.
4. Version and score your top prompts by tier. So a re-route is a decision you can check, not a hope.
5. Stop hard-coding the number, hard-code the tier. Your library should say "this prompt runs on Terra," so that when 5.7 ships, nothing in your prompts has to change.

If you want the scoring and versioning layer in place before you start moving prompts between tiers, Keep My Prompts scores every prompt you save and keeps a version history, so a tier change is a measured move with a paper trail. Free to start.

The shape is the story

OpenAI did not just ship a better model on July 9. It shipped a grid of prices and effort levels and made the tiers durable, which quietly turns model selection into a routing problem you solve once and keep. The teams that come out ahead are not the ones who switched their string to gpt-5.6 fastest. They are the ones whose prompt library already knew which tier each prompt belonged on.