OpenAI ships GPT-5.6 as a three-model family: Luna, Terra, and Sol
GPT-5.6 hit general availability July 9 in three sizes, priced $1 to $5 per million input tokens, with a 1M context and tool calling that writes its own orchestration code.
OpenAI’s GPT-5.6 went generally available on July 9, and it is not one model, it is three: Luna, Terra, and Sol, smallest to largest, priced per million tokens at $1/$6, $2.50/$15, and $5/$30. For comparison, Claude Opus runs $5/$25 and Claude Fable 5 runs $10/$50. All three carry a February 16, 2026 knowledge cutoff, a 1M-token context window, and 128K max output.
The capability story is agentic. Programmatic tool calling lets the models write JavaScript that orchestrates their own tool calls, they can spin up subagents for parallel work, and there are explicit prompt-cache breakpoints. The benchmark picture, all of it attributed: per Artificial Analysis, Sol scores 59 on the Intelligence Index, one point behind Fable 5 at 60, at roughly a third the per-task cost, and it tops the Coding Agent Index at 80 inside OpenAI’s Codex environment. Per Willison, Sol sets a new high on Agents’ Last Exam at 53.6, thirteen points clear of Fable 5.
None of this runs on your bench, and that is the part worth sitting with. When frontier gets this cheap per task, “open weights are cheaper” stops being an argument. What is left is the argument that was always better anyway: weights on your own disk answer to you, work offline, and never change under your feet. The price war just made the honest case for local the only case.