OpenAI launched GPT-5.6 on July 9 as a three-model family: Sol, Terra, and Luna.
The launch puts GPT-5.6 into ChatGPT Work, Codex, and the OpenAI API. OpenAI says Sol is the highest-performing model, Terra is the default for most work, and Luna is the smallest and fastest option for lower-cost tasks.
API pricing follows that split. OpenAI lists Sol at $5 per million input tokens and $30 per million output tokens. Terra is $2.50 input and $15 output. Luna is $1 input and $6 output. Cached input is priced lower for all three.
The model launch is also an agent launch
The new API surface matters as much as the model names. OpenAI says the Responses API now supports Programmatic Tool Calling, which lets developers expose typed functions to the model without forcing the model to produce raw JSON arguments in text.
OpenAI is also opening a multi-agent beta. In the launch description, the primary model can delegate subtasks to subagents with narrower instructions, tools, and context. That is a direct product bet on agent decomposition rather than one giant prompt doing all the work.
The Codex angle is similarly practical. OpenAI says GPT-5.6 improves on software-engineering tasks and is available in Codex, where long tasks, tool calls, and repository context make cost and reliability visible quickly.
Sol, Terra, and Luna make pricing a product decision
Three model tiers give teams a more explicit routing problem. Sol is the model to test for hard reasoning, migration work, and high-stakes agent runs. Terra is likely the default candidate for everyday coding and analysis. Luna is the candidate for support, extraction, light tool use, and workflows where latency or budget dominates.
That is not a purely financial choice. A cheaper model that needs more retries can cost more in real work than a more expensive model that finishes cleanly. Teams should measure accepted outputs, tool calls, wall-clock time, cached-token reuse, and human review edits rather than comparing per-token prices alone.
OpenAI’s system card gives the other half of the decision. It classifies GPT-5.6 under OpenAI’s Preparedness Framework and discusses cyber, biological, persuasion, and autonomy evaluations. Those details do not prove safety in deployment, but they give enterprise and governance teams a concrete artifact to review before broad rollout.
