OpenAI made the GPT-5.6 family generally available on July 9 after a limited preview. Sol is the flagship tier, Terra is positioned as a balance of performance and cost, and Luna is the fastest and least expensive option. The family is available across ChatGPT, Codex, and the API, so descriptions of GPT-5.6 as preview-only are now out of date.

The release is built around efficiency as much as headline capability. OpenAI argues that stronger models can cost less per successful task when they finish work in fewer attempts and produce fewer tokens. It also introduces an ultra effort setting for Sol that coordinates multiple agents across parallel workstreams for demanding tasks.

What changed for users and developers

  • The three-tier family gives teams a routing choice instead of a single flagship model.
  • OpenAI reports gains across coding, computer use, design, knowledge work, science, and cybersecurity.
  • Long-context evaluations extend to very large inputs, but real retrieval quality can vary by document structure and task.
  • The most sensitive defensive cyber capabilities use a trusted-access path with identity and account-security requirements.

OpenAI’s published benchmark tables are extensive, but they remain vendor evaluations with harness, latency, and cost assumptions. The company explicitly notes that some latency and API-cost figures are simulated and that real-world results can vary. Buyers should therefore use the launch numbers to form test hypotheses, not to skip their own evaluation.

Safety is part of the product boundary

OpenAI says GPT-5.6 is more capable in cybersecurity and biology than earlier models but does not cross its Critical threshold in either domain. For cyber work, default safeguards are more conservative, while verified defenders can request broader capability through OpenAI’s trusted-access program. That may help legitimate security teams, but it also adds account, identity, and workflow dependencies that should be evaluated before migration.

The company says it used extensive red teaming before general availability and added a reasoning monitor alongside model protections, real-time checks, continuous monitoring, and account enforcement. Those are meaningful design claims. Independent assessment will be needed to understand false positives, adaptive attacks, and whether the controls preserve routine defensive work.

For production teams, the sensible comparison is not Sol versus a competitor on one benchmark. Test a representative task set across Sol, Terra, Luna, the current incumbent, and a fallback. Measure acceptance rate, review time, latency, tool reliability, cost, and failure severity. General availability makes that evaluation possible; it does not make the answer universal.