GPT-5.6 Sol, Terra and Luna: how the new family is positioned

What the tier names actually mean

Model naming has been chaotic for years, and this scheme is a deliberate response to it. OpenAI’s framing separates two things that used to be tangled together: which generation of technology you are using, and how much capability you are paying for.

The consequence is worth thinking through, because it changes how you should write things down. Under a single-version scheme, “we use the newest model” was an unambiguous instruction. Under a tiered one it is not, because GPT-5.6 Sol, Terra and Luna are all the newest model — they differ in tier, not recency.

Three practical habits follow. Record the full tier name in any evaluation, not just the version number, or your results will be unreproducible. Expect tiers to advance independently, so a Terra-level model in a future generation may exceed today’s Sol. And when a vendor quotes a benchmark, check which tier it refers to, since a family-level headline usually describes the flagship only.

OpenAI also describes an ultra setting for demanding work that coordinates multiple agents across parallel workstreams — a third axis beyond generation and tier, and another parameter to record.

OpenAI made the GPT-5.6 family generally available on July 9, 2026. The family has three named tiers: Sol as the flagship, Terra as the balanced everyday model and Luna as the fastest, lowest-cost option. OpenAI also describes an ultra setting for demanding work that coordinates multiple agents across parallel workstreams.

The practical meaning of three tiers

A tiered family lets teams match model expense to task difficulty, which is the whole reason GPT-5.6 Sol, Terra and Luna exist as separate names. High-volume classification, extraction or routine drafting may not need the flagship. Complex software changes, research synthesis or multi-step tool use may benefit from a more capable tier if it reduces failures and rework.

OpenAI reports performance and cost advantages for GPT-5.6 on its chosen evaluations. These are company-reported results. They should inform a shortlist, not replace a test using your prompts, tools and acceptance criteria.

Published pricing, and blended cost

OpenAI publishes list rates per million tokens, and the spread across the family is what makes tier routing worth engineering.

The interesting number is not any single row but the blend. Suppose a workload routes 80% of requests to Luna and 20% to Sol. The effective output rate is (0.8 × $6) + (0.2 × $30) = $4.80 + $6.00 = $10.80 per million output tokens — about a third of what routing everything to Sol would cost, while the hard 20% still gets flagship treatment.

That arithmetic is the entire commercial argument for a tiered family, and it is why the engineering work worth doing is classification: deciding reliably which requests are hard. A router that misclassifies difficult work down a tier converts a cost saving into a rework bill, so measure the failure rate of the router itself, not just of the models.

A useful starting rule: route to the cheapest tier by default, escalate on a detectable failure signal — a failed test, a schema validation error, a low-confidence extraction — and log every escalation so you can see what the classifier is missing.

The benchmark claims, attributed

OpenAI’s published figures for the family, quoted as the company’s own results rather than as independent findings: Sol “sets a new high of 53.6” on Agents’ Last Exam, described as “eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points”; on coding, “80, 2.8 points above Fable 5, while using less than half the output tokens”; and on the ExploitBench security evaluation, “73.5% versus GPT-5.5’s 47.9%.”

Two observations rather than a verdict. The token-efficiency claim matters commercially as much as the score, because halving output tokens partially offsets a higher per-token rate — a three-times rate applied to half the tokens is roughly 1.5 times the cost. And the comparison names a specific competitor configuration, “adaptive reasoning,” which is a reminder that cross-vendor benchmarks are conditioned on settings that a headline number does not carry.

Neither observation tells you whether GPT-5.6 Sol, Terra and Luna suit your workload. Only your own task set does that.

A disciplined selection process

1. Create a representative set of tasks and expected outputs.
2. Define pass, partial-pass and fail criteria before testing.
3. Run each task more than once at fixed settings.
4. Record tokens, latency, tool calls, human correction and total cost.
5. Choose the lowest tier that meets the required success rate.

For agentic workflows, also measure safe stopping. A system that recognizes missing permission or conflicting evidence can be more valuable than one that confidently continues. Keep external actions reversible where possible, and review logs for unexpected data access.

Watch pricing and availability

OpenAI’s preview announcement included per-token pricing and prompt-caching terms, while the general-availability post describes broader access across products. Commercial terms can change. Verify the current API model page and account availability before calculating a production budget, since GPT-5.6 Sol, Terra and Luna are documented as tiers that advance independently of the version number.

Sources and methodology

This explainer uses OpenAI’s preview, launch and engineering posts. It does not independently validate OpenAI’s benchmark comparisons. RankBoast recommends testing complete workflows because model-only benchmark scores omit integration and review costs. Pricing was re-verified against OpenAI’s published rates in August 2026.

• OpenAI official source
• OpenAI official source
• OpenAI official source