GPT‑5.6 Family Debuts, Then OpenAI Cuts Luna's Price 80%

OpenAI released the GPT-5.6 family on July 9, 2026, following a limited preview: Sol, the flagship; Terra, a balanced model for everyday work; and Luna, the most cost-efficient of the three. It’s the first time this site has covered an OpenAI model release. Three weeks later, on July 30, OpenAI cut the price of two of them: Luna by 80%, Terra by 20%, Sol unchanged.

Meet Sol, Terra, and Luna

The three models are GA across ChatGPT, Codex, and the API. Sol is the flagship, tuned for the hardest agentic and coding work. Terra is positioned as the everyday, balanced pick. Luna is the cost-efficient option, meant to make high-volume agentic work affordable rather than serving as a stripped-down fallback.

OpenAI’s own framing for Sol is efficiency alongside capability: “stronger performance per dollar” is the phrase it uses across the release, more successful work for the same spend rather than just a bigger number on a leaderboard.

A Price Cut Three Weeks After Launch

The July 30 update reads plainly: “OpenAI reduced the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%.” Sol’s price stayed exactly where it launched. Exact per-token dollar figures for a model that just had its pricing rewritten are worth checking at the source rather than trusting a number printed here or on any aggregator, since the pre-cut figures that were still circulating days after the change were off by as much as 80%. Budget from OpenAI’s current pricing page at the time you’re actually spending, not from a number written into an article weeks or months earlier.

The Benchmarks: Agents’ Last Exam and the Coding Agent Index

On Agents’ Last Exam, a test of long-running professional workflows across 55 fields, OpenAI says Sol scores 53.6, “eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points.” On the Artificial Analysis Coding Agent Index, OpenAI says Sol at max reasoning hits 80, “2.8 points above Fable 5, while using less than half the output tokens.”

That second comparison is the one worth sitting with: a 2.8-point lead paired with under half the output tokens is a cost story as much as a capability one, and OpenAI applies that same efficiency-first framing to Terra and Luna too, not only to the flagship.

The Reasoning Ladder and Programmatic Tool Calling

GPT-5.6’s reasoning-effort ladder now runs xhigh → max → ultra. max gives the model more time than xhigh to reason, explore alternatives, and revise its approach; ultra goes further and coordinates four agents in parallel by default, trading higher token use for stronger results and faster time-to-result on demanding tasks. Developers can build similar setups themselves through the new multi-agent beta in the Responses API.

The other developer-facing piece is Programmatic Tool Calling: instead of round-tripping every tool response back through the model, GPT-5.6 can write and run lightweight code that coordinates tools, filters intermediate results, and decides the next action as the work unfolds. For tool-heavy agent tasks, that’s fewer tokens and fewer model round trips for the same job, a direct lever on cost independent of which model tier you’re running.

What It Means for Developers

This is a real introduction to a family, not a footnote to a price cut: Sol, Terra, and Luna give OpenAI a three-tier lineup aimed squarely at the same agentic and coding work Anthropic’s Claude Opus 5 and Claude Sonnet 5 are built for. If you’re running any of these models through an automated Codex review in a CI pipeline, the same rule applies there as everywhere else in this piece: check the live rate card before you budget, because OpenAI has shown it will move prices weeks after a launch, not years.