OpenAI paper lists three GPT-5.6 Pro models

A new OpenAI benchmark table now shows three distinct Pro variants for GPT-5.6: Luna Pro, Terra Pro, and Sol Pro. This marks the first time the Pro tier has been split into multiple options rather than offering a single top-tier model.

OpenAI officially announced the GPT-5.6 generation in late June, dividing it into Sol for difficult tasks, Terra for high-volume business workloads, and Luna for faster, cheaper everyday queries. The Pro variants were not part of that initial announcement.

Instead, a separate paper focusing on genomics benchmarks revealed these Pro models. The results table includes rows for “GPT-5.6 Luna Pro,” “Terra Pro,” and “Sol Pro,” each marked as “Pro (Extended)” runs.

Three variants replace the single best model

Under the current structure, ChatGPT Pro was simply the single best model available, sitting one tier above everything else. The new paper suggests this is changing. It lists three parallel Pro variants that mirror the standard GPT-5.6 lineup: a fast one, a high-volume one, and a max-performance one.

Comparing each standard tier at its highest reasoning setting against its Pro variant shows how the gains play out. All values are pass rates on the full 129-task suite:

In this case, the Pro boost shrinks as you move up the ladder. Luna Pro gains a full seven points over its standard version, while Sol Pro picks up less than three. Extra compute lifts weaker tiers more: Terra Pro lands at 28.5 percent, nearly matching standard Sol at 28.7 percent. This means a high-volume Pro variant performs almost as well as the best standard flagship.

A shift in how Pro works

This split would be the first major change to the Pro offering since ChatGPT Pro launched. Instead of one expensive top tier, Pro could become its own three-model lineup where users pick between speed, throughput, and maximum reasoning power based on the task at hand.

Whether this tiered structure will actually show up in ChatGPT is not clear from the paper. The names come only from the benchmark table so far.

One detail stays hidden, too. For the standard GPT models, the paper reports average token usage as a rough proxy for compute cost, about 33,200 tokens for Sol at its highest setting. For the Pro runs, that number is missing. The authors say no comparable token accounting was available, but the more likely explanation is that OpenAI simply does not want to share those figures.