What happened
OpenAI released the GPT-5.6 model family for general availability: Sol, the new flagship, Terra, a balanced model for everyday work, and Luna, the cheapest and fastest of the three. OpenAI says the family delivers more capability per token, with Sol setting a new high score on Agents’ Last Exam, a benchmark of long-running professional tasks across 55 fields.
The launch also introduces ultra, a new setting that coordinates multiple agents across parallel workstreams to finish complex, long-running work faster. OpenAI is pitching Sol specifically on stronger computer use and design judgment, meaning it can inspect, refine, and hand back work that’s closer to done without another round of edits.
This article is the introduction to the model family. OpenAI later changed launch defaults and restored allowances after users reported friction. Further usage resets and capacity changes show why the surrounding product matters as much as the model name.
Why it matters
The three tiers create a clearer capability and cost decision for builders using GPT-5.6 directly or through a coding tool. Sol is the high-reasoning option, Terra is the everyday choice, and Luna is aimed at faster, cheaper tasks.
The more useful question is not whether Sol leads a benchmark. We flagged yesterday that OpenAI’s own benchmark audit means headline scores like Sol’s Agents’ Last Exam result deserve a healthy amount of skepticism. A model claiming to beat the field on a benchmark is not the same as it beating the field on your codebase.
Who should care
Builders choosing a model for difficult reasoning, everyday coding, or quick iterations should pay attention. Teams evaluating AI development tools should also test the models against their own code and workflow rather than relying on benchmark scores alone.
What builders should do next
Start with Terra on a normal task if it is available in your chosen tool. It is positioned as the everyday default. Save Sol for something that is actually stalling your current model: a large refactor, a difficult multi-file bug, or a task that needs several coordinated steps.
Use Luna for quick, lower-cost iterations and compare the result against Terra on the same task. Track quality, latency, and usage cost on your own codebase before making Sol the default for regular work.
The practical takeaway is simple: choose the smallest GPT-5.6 tier that handles the task, then reserve Sol for work that benefits from its higher reasoning ceiling.
End of article