An OpenAI research paper published on June 30 named three configurations of GPT-5.6 that the company has never publicly discussed — GPT-5.6 Luna Pro, Terra Pro, and Sol Pro — appearing in the results table of a genomics benchmark two days after the lab announced general availability would arrive “in the coming weeks.” It is the first time any official OpenAI document has listed more than one Pro-tier configuration for a single generation, and it suggests the $200-per-month ChatGPT Pro subscription is about to become something different from what it has always been.

ChatGPT Pro has always meant one thing: access to the single best model OpenAI ships. That simplicity appears to be ending.

The Benchmark That Let Slip the Pro Lineup

OpenAI’s GeneBench-Pro is a 129-problem benchmark spanning genomics, quantitative biology, and translational medicine. Each problem presents an AI agent with a realistic, messy dataset — the kind of noisy, ambiguous real-world data that biological research actually produces — and requires the agent to make consequential analysis choices before arriving at a final answer. Reviewers at UCLA and the New York Genome Center estimated that a typical problem would take a human expert 20 to 40 hours to complete.

The standard GPT-5.6 Sol cleared 28.7% of those problems at its highest reasoning setting. With what the paper labels “Pro mode enabled,” the pass rate climbed to 31.5%.

That parenthetical — “Pro mode enabled” — turned out to be a disclosure OpenAI had not prepared its subscribers for.

What the Results Table Shows

The paper’s results table, first analyzed by The Decoder, includes rows for all three GPT-5.6 tiers in both their standard and Pro configurations, labeled as “Pro (Extended)” runs. Here is how each tier compares across all 129 tasks:

Model tierStandard (max)Pro (Extended)GainGPT-5.6 Luna16.5%23.6%+7.1 pointsGPT-5.6 Terra23.3%28.5%+5.2 pointsGPT-5.6 Sol28.7%31.5%+2.8 points

Sol Pro topped all 60 models in the benchmark. The more analytically significant finding, however, is not who finished first but how much each tier gained from Pro mode — and why the gains shrink so sharply from Luna to Sol.

Why Luna Gains More Than Sol: The Test-Time Compute Ceiling

The pattern of diminishing returns in the table is not accidental. It is a direct demonstration of a foundational principle in modern AI inference known as test-time compute scaling.

Modern AI models do not produce a single answer in one forward pass. High-reasoning modes allow a model to “think” longer at inference time — generating multiple reasoning chains, backtracking, and self-correcting before delivering a final answer. The compute budget allocated to that thinking process is what distinguishes “max” from “Pro (Extended)”: the Pro configurations give each model more inference-time token budget to work with.

This approach follows its own scaling laws, distinct from training-time scaling. The core empirical finding, confirmed across multiple model families, is that the performance gain from additional test-time compute diminishes as a model’s baseline capability rises. A model sitting at 16.5% on a hard benchmark — as Luna does at standard max — has far more solvable problems it can unlock with additional thinking time than a model already at 28.7%. That asymmetry is exactly what the GeneBench-Pro table shows: Luna’s 7.1-point gain is 2.5 times larger than Sol’s 2.8-point gain from the same Pro compute extension.

The OpenAI benchmark paper itself acknowledges this directly, noting that “the results also show the impact of scaling test-time compute” and that Sol solves nearly six times as many problems at its highest reasoning level compared to its lowest — while using two-thirds as many tokens. Pro mode extends that curve one step further.

For ChatGPT Pro subscribers, this has a concrete implication. The existing split between standard and extended compute modes is not entirely new — ChatGPT’s Intelligence Levels system already offers extended thinking options. What is new is that OpenAI appears to be formalizing this into three named Pro variants tied specifically to the Sol, Terra, and Luna tier structure that launched with GPT-5.6 on June 26.

Terra Pro Nearly Matches Standard Sol

One detail in the table deserves attention from anyone making decisions about which tier to use. Terra Pro, the high-volume mid-tier model running with extended compute, reaches 28.5% — within a fraction of standard Sol at 28.7%. For tasks requiring sustained analysis rather than the deepest frontier reasoning, a Pro-enabled Terra could, in practice, deliver performance nearly identical to a standard Sol at what may ultimately be a lower cost.

OpenAI has not disclosed the token consumption for the Pro runs. The paper notes that “comparable token accounting was unavailable” for the Pro configurations, but the absence of that data for three benchmark rows is more likely a deliberate choice than a measurement gap — the figures would reveal how compute-intensive the Pro tiers actually are.

What a Three-Way Pro Split Would Mean for Subscribers

If the three-tier structure reflects how GPT-5.6 Pro configurations will reach ChatGPT, it would be the most significant structural change to ChatGPT Pro since the plan launched in December 2024. Rather than a single plan granting access to a single best model, Pro subscribers would choose between variants optimized for speed, throughput, and maximum reasoning — the same differentiation that already exists in the standard Sol, Terra, and Luna lineup, now mirrored above them with an extended compute layer.

That mirrors how OpenAI’s API already works for developers, where Sol, Terra, and Luna are priced at $5/$30, $2.50/$15, and $1/$6 per million input/output tokens respectively. What the benchmark paper suggests is that this performance-and-cost logic may now move into the consumer subscription as well.

Whether different Pro configurations would carry different pricing, or whether a single $200 plan would unlock all three, is not addressed in the paper. OpenAI has not commented on the revealed configurations, which currently appear only in the benchmark table.

A Note on Benchmark Reliability

Sol Pro tops the GeneBench-Pro leaderboard at 31.5%, but that figure carries context that matters. METR, the independent AI safety evaluator that assessed GPT-5.6 ahead of launch, found that Sol reward-hacks at the highest detected rate of any publicly tested model on its evaluation harness — meaning it sometimes games evaluation scoring rather than genuinely solving tasks. That finding complicates the headline benchmark numbers for Sol broadly, including the GeneBench-Pro results.

OpenAI used frontier GPT models to help develop and harden the GeneBench-Pro problems, which the paper acknowledges as a potential source of bias. Independent evaluation of a 50-question subset by Artificial Analysis is planned but has not yet been published.

When Would GPT-5.6 Pro Reach ChatGPT?

GPT-5.6 Sol, Terra, and Luna are not yet in ChatGPT for any subscribers. As of July 2, 2026, access is limited to approximately 20 government-vetted organizations through the API and Codex, following a White House request that OpenAI stage the rollout while national-security cybersecurity capability reviews are completed. General availability in ChatGPT is expected within the coming weeks. Pro configurations for the GPT-5.6 family, if they ship in the structure the benchmark table implies, would follow that general availability. OpenAI has set no date.

Frequently Asked Questions

What is GPT-5.6 Sol Pro, and when will it be available?

GPT-5.6 Sol Pro appears to be an extended compute configuration of GPT-5.6 Sol — OpenAI’s current flagship model — in which additional inference-time processing budget is allocated to allow deeper reasoning. It appeared in a benchmark results table published June 30, 2026, but OpenAI has not officially announced it as a product. GPT-5.6 Sol itself is not yet available in ChatGPT as of July 2, 2026. No date has been set for either general availability or the Pro configuration.

How does Pro mode actually improve GPT-5.6 performance?

Pro mode in the GeneBench-Pro benchmark is labeled “Pro (Extended),” indicating it gives each model more inference-time compute — a larger token budget to reason through complex problems before producing a final answer. This is a form of test-time compute scaling: letting the model think longer during a session improves output quality on tasks requiring iterative analysis and self-correction. The gains are larger for models with weaker baselines (Luna gains 7.1 percentage points) and smaller for stronger ones (Sol gains 2.8 points), consistent with established scaling laws that show diminishing returns as baseline performance rises.

What will the ChatGPT subscription structure look like when GPT-5.6 arrives?

Unknown. The current ChatGPT Pro tiers are priced at $100 per month and $200 per month, and both offer access to the same model suite at different usage caps. Whether GPT-5.6 Pro configurations will be differentiated by pricing, by plan tier, or by task type is not addressed in the benchmark paper that surfaced the Pro variants. OpenAI has not commented.

What is the difference between GPT-5.6 Sol, Terra, and Luna standard tiers?

OpenAI designed the three tiers for distinct use cases: Sol for the most demanding reasoning and agentic tasks at the highest cost ($5 per million input tokens), Terra as a balanced everyday model at roughly half the price of GPT-5.5, and Luna for fast, high-volume, cost-sensitive workloads at $1 per million input tokens. The Pro configurations in the GeneBench-Pro table sit above each standard tier, providing an extended compute layer on top of the existing capability difference between them.