OpenAI GPT-5.6: Luna, Terra and Sol Pro variants revealed

OpenAI's genomics benchmark paper embeds three Pro models for GPT-5.6 named Luna Pro, Terra Pro and Sol Pro, appearing in a results table as "Pro (Extended)" runs. The paper shows Sol Pro hitting a 31.5 percent pass rate on the 129-task suite, higher than standard Sol at 28.7 percent and beating the best non-GPT score, Claude Opus 4.8 at 16.0 percent.

OpenAI officially unveiled the GPT-5.6 generation in late June with three standard tiers: Luna for faster, cheaper queries, Terra for high-volume business workloads and Sol for the hardest tasks. The paper adds Pro variants for each tier, though it does not say whether those Pro models will ship inside ChatGPT.

What did the paper disclose about Pro variants?

The paper lists Luna Pro, Terra Pro and Sol Pro as "Pro (Extended)" runs in its benchmark tables, marking the first time Pro appears as three parallel variants rather than a single top tier. The standard GPT-5.6 lineup was already split into Luna, Terra and Sol, and the Pro entries mirror that split: a fast variant, a high-volume variant and a max-performance variant.

The names appear only in the benchmark table. The paper does not confirm a ChatGPT rollout or any product changes, and it omits token usage for the Pro runs that it provides for standard models.

How large are the performance gains from Pro?

Pro runs improve pass rates across all three tiers, with larger lifts on weaker tiers and smaller gains at the top end. On the full 129-task suite the paper lists these exact pass rates and gaps: Luna standard at 16.5 percent versus Luna Pro at 23.6 percent, a gain of 7.1 points; Terra standard at 23.3 percent versus Terra Pro at 28.5 percent, a gain of 5.2 points; Sol standard at 28.7 percent versus Sol Pro at 31.5 percent, a gain of 2.8 points.

The paper highlights that Terra Pro at 28.5 percent nearly matches standard Sol at 28.7 percent, meaning a high-volume Pro variant performs almost as well as the best standard flagship in this genomics benchmark. The benchmark measures pass rate as how often a model completes the full multi-step analysis without errors and arrives at the correct final answer.

What does the paper hide about Pro?

The paper reports average token usage for standard models, for example about 33,200 tokens for Sol at its highest setting, but it does not report comparable token accounting for the Pro runs. The authors say no comparable token accounting was available; the article notes a more likely explanation is that OpenAI chose not to share those figures.

That missing token data leaves open questions about the compute cost and billing implications of Pro (Extended) runs, and whether Pro variants will carry distinct pricing or usage limits if they appear in a product.

Why it matters

If OpenAI turns Pro into a three-model lineup, Pro would stop being a single pricey top tier and become a choice between speed, throughput and peak reasoning power. That would change how organizations and power users select a paid tier: some workloads might be cheaper on Terra Pro if it matches flagship standard performance while offering higher throughput, while others may need Sol Pro for the highest pass rates on complex multi-step tasks.

The missing Pro token accounting matters for buyers. Without those numbers it is impossible to compare cost-per-task across standard and Pro runs, or to predict how Pro will affect billing and capacity planning for heavy users.

What to watch

Watch for OpenAI to surface these Pro variants in product listings or pricing pages and for any subsequent papers or documentation that disclose token usage for Pro runs. A public change to ChatGPT Pro that mirrors the papers three-way split, or a follow-up benchmark that adds Pro token accounting, would confirm whether these entries are experimental benchmarks or a near-term product shift.