GPT-5.6 Sol, Terra and Luna: which model is good at what?

OpenAI introduced the GPT-5.6 family as Sol, Terra and Luna. The launch began as a limited preview; the Help Center page updated on July 10 now says Sol is rolling out gradually to eligible ChatGPT plans. The important distinction is that Sol is coming to standard ChatGPT conversations, while Terra and Luna are not available in the standard model picker. Those two are positioned across Work, Codex and the API.

It is more useful to read them as three work priorities than as three personalities of the same model: Sol for capability and deep reasoning, Terra for balance, Luna for speed and lower cost. In my earlier GPT-5.6 note, I argued that the real jump might be reliability across long agentic tasks rather than one polished answer. The official launch largely points in that direction.

Sol: the flagship for difficult, long-running work

GPT-5.6 Sol is the most capable model in the family. OpenAI positions it for complex work across software engineering, professional knowledge work, research, science, computer use, design and cybersecurity. In ChatGPT, Medium, High and Extra High reasoning use Sol, while the Pro option for the hardest workflows is powered by Sol Pro.

My practical choice would be Sol for multi-file refactors, hard-to-diagnose bugs, research that must reconcile several sources, long tool-using tasks, or decisions where a wrong answer is expensive. For a simple summary or a routine operation that is easy to verify, Sol may add unnecessary latency and resource use.

Terra: the balanced option for everyday production

GPT-5.6 Terra balances capability, speed and cost. OpenAI describes it as the model for everyday work. It is not selectable in standard ChatGPT conversations; depending on plan and product access, it is available through Work, Codex and the API.

My practical inference from that positioning: routine code changes, document review, repeatable office workflows, tasks with clear acceptance criteria, and systems that run many subagents are Terra's natural territory. When every step does not need Sol-level depth, Terra may be the sensible middle path for a smoother production line without giving up too much capability.

Luna: when speed and volume matter

GPT-5.6 Luna is the fastest and lowest-cost model in the family. It also does not appear in the standard ChatGPT model picker; it is used through Work, Codex and the API.

The use cases here are also my inference: high-volume classification, first-pass summaries, quick code scans, data cleanup, draft generation and parallel subtasks whose outputs are easy to check. A vague, multi-step problem should not be handed to Luna just because it is fast. Speed is valuable when there is a reliable way to catch wrong results.

Where did GPT-5.6 actually improve?

OpenAI's first published evaluations focus heavily on agentic work. Sol sets a new state of the art on Terminal-Bench 2.1, which measures command-line tasks that require planning, iteration and tool coordination. On GeneBench v1, covering long-horizon genomics and quantitative-biology workflows, it outperforms GPT-5.5 while using fewer tokens. In cybersecurity, the emphasis is on vulnerability research, code review, patch development and defensive testing.

A new max reasoning effort gives Sol more time on difficult problems. Ultra mode uses subagents to go beyond the capability of a single agent. Together they show that GPT-5.6 is not merely a chat model that writes better: the target is a system that plans work, uses tools and completes dependent steps.

We can already see that shift in real usage. My note on Codex usage data showed people using agents not only to generate code, but to delegate tasks and run workflows. The Sol, Terra and Luna split tries to scale that behavior by task type instead of forcing every job through one model.

Which one will appear in ChatGPT?

• Instant: GPT-5.5 Instant remains the default for fast, everyday responses.

• Medium, High and Extra High: Different GPT-5.6 Sol reasoning levels on eligible plans.

• Pro: GPT-5.6 Sol Pro for difficult and longer-running workflows.

• Terra and Luna: Not selectable in standard ChatGPT conversations; availability across Work, Codex and the API depends on plan and product access.

The rollout is gradual, so it is normal not to see Sol immediately. Plan availability and usage limits can also change; OpenAI's current Help Center page is the right place to check access.

How I would choose

• Sol: When mistakes are expensive, the task has many dependent steps, or deep reasoning and tools are required.

• Terra: When the everyday production task is clear and the capability-speed balance matters.

• Luna: When speed and volume come first and the result can be verified cheaply by automation or a person.

• Instant: For conversation, short explanations and simple daily questions that do not need heavy reasoning.

It is healthier to look at task risk before the model name. My coding-agent guide for solo builders follows the same principle: split the work and give each piece a model that is sufficient, rather than keeping the strongest option on for everything.

More capability should not mean less control

The GPT-5.6 system card contains an important warning. In agentic coding evaluations, the family showed a greater tendency than GPT-5.5 to go beyond the user's intent, although OpenAI says absolute rates remained low. Sol and Terra could find vulnerabilities and pieces of exploits, but did not complete autonomous end-to-end attacks against hardened targets in the reported tests.

The practical implication is simple: a stronger agent does not justify weaker supervision. Destructive file operations, production writes, sending email, spending money or changing access settings still need explicit confirmation; code changes need diffs and tests; external systems need least-privilege access. The boundary for irreversible action should remain with the user.

Bottom line

The real message behind Sol, Terra and Luna is larger than three new names. Instead of one universal 'best model,' OpenAI is asking users to choose among depth, balance and speed within the same family. My short version: reason with Sol, produce with Terra, scale with Luna. In all three cases, the process that verifies the result matters more than the model label.