GPT-5.6 Sol, Terra and Luna: what changes for developers

OpenAI introduced the GPT-5.6 family with three names that, for the first time, try to explain the model's role before the user checks a technical table: Sol, Terra and Luna.

Sol represents maximum capability; Terra aims for balance; Luna prioritizes speed and cost. The announcement was a limited preview, not an open release for everyone.

The simplest comparison is transportation. You wouldn't use a semi-truck to go buy a loaf of bread, but you wouldn't move a house with a motorcycle either.

The same happens with artificial intelligence: the most powerful model isn't always the best choice. Sometimes answering quickly, spending less or processing thousands of repetitive tasks matters more.

Not one model, but three tiers

OpenAI is debuting a structure that could be easier to understand than the earlier collection of names and versions.

The number 5.6 identifies the generation, while Sol, Terra and Luna work as permanent capability tiers. Those names could stay even if each tier advances at different paces.

Official prices during the preview.

GPT-5.6 Sol is the flagship model. It's designed for complex problems, long processes and jobs where quality weighs more than cost.

GPT-5.6 Terra is the balanced option. OpenAI claims it offers performance competitive with GPT-5.5 at roughly half the price.

GPT-5.6 Luna is the fast, affordable model. Its goal isn't to win every test, but to handle tasks where speed and volume matter more than squeezing out the last percentage of reasoning.

Sol: programming beyond completing lines

Long jobs and tool coordination

The interesting part for development isn't that Sol can write a function. Today's models already do that reasonably well.

The advance is in sustaining long tasks: reviewing a project, using tools, detecting that an attempt failed, correcting course and continuing.

OpenAI reports that Sol reached a new high on Terminal-Bench 2.1, an evaluation that puts the model in front of real terminal tasks requiring planning, iteration and coordination.

It's like evaluating a mechanic not on identifying a wrench, but on handing them a car with an unclear fault and seeing whether they manage to diagnose it, try solutions and leave it running.

There's also a new reasoning level called max, which lets you devote more time to a problem.

The ultra mode goes a step further: it uses subagents, something like splitting a project among several coordinated specialists instead of asking everything of one person.

Cybersecurity: more capability, but not a license to attack

OpenAI describes Sol as its most capable model for cybersecurity.

On ExploitBench, a test focused on vulnerability research and exploitation, it achieved performance competitive with Mythos Preview while using about a third of the output tokens.

All three models also show improvements on ExploitGym as they are given more reasoning time.

This doesn't mean OpenAI is releasing an “automatic hacker.”

According to its own evaluations, GPT-5.6 is better at finding and fixing vulnerabilities than at carrying out complete attacks on real systems.

Sol identified bugs and the pieces needed to build an exploit, but it did not autonomously produce a complete working chain in the Chromium and Firefox tests. That's why it was not classified at the Cyber Critical level.

The difference matters.

Finding a poorly closed door and helping repair it isn't the same as walking in, going through the building and switching off the alarms.

Defensive capability can be used to review software, prepare fixes and strengthen systems, although it also calls for stricter controls.

OpenAI added generation-time filters, account-level signals, additional reviews and specific monitoring for sensitive requests.

It also devoted more than 700,000 A100 GPU-equivalent hours to automated testing against attempts to bypass its controls.

Terra and Luna may be more useful for most people

Sol grabs the headlines, but Terra and Luna will probably have more everyday impact.

In a company, the price difference can decide which model is used in each process:

• Sol could be reserved for reviewing an architecture, investigating a complex failure or coordinating a big change.
• Terra would fit daily analysis, development and documentation.
• Luna would serve to classify information, summarize requests or handle large volumes with fast responses.

It's the same logic as an enterprise cloud: not every process needs the biggest machine.

Choosing correctly can matter as much as having access to the most advanced model.

Still not available to everyone

The GPT-5.6 family is in a limited preview for a small group of selected organizations.

During this stage it is only available through the API and Codex for approved participants; it is not enabled in ChatGPT.

OpenAI says it plans to expand access to ChatGPT, Codex and the API in the coming weeks, but has not published a firm date.

It also announced that Sol will run on Cerebras infrastructure at speeds of up to 750 tokens per second during July, initially for selected customers.

This forces us to put the results in context: they are preliminary figures published by OpenAI during a preview.

The company said it will share a broader evaluation when general availability arrives. It's wise to wait for independent testing and real-world experience before treating GPT-5.6 as an absolute winner.

The essentials

GPT-5.6 doesn't just increase capability. It also organizes the lineup around a more practical decision:

• Sol: maximum intelligence for hard problems and long processes.
• Terra: balance between capability and budget.
• Luna: speed and cost for high-volume tasks.

For programming, the most important signal is the move toward agents able to sustain work for longer.

For cybersecurity, the value is in detecting and fixing flaws, accompanied by controls that try to limit offensive use.

Conclusion: choosing the model will be part of designing the system

The question will no longer be only “which is the smartest model?”

We'll also have to decide how much each answer is worth, how long the user can wait and what level of capability the job really requires.

GPT-5.6 Sol shows where the frontier is heading. Terra and Luna show how OpenAI wants to turn that frontier into usable products.

For those building software, the most sensible strategy will be to test each task with the cheapest model that meets the expected result and reserve maximum power for the problems that truly need it.