GPT-5.6 family lands with Luna, Terra, Sol

Before, users relied on GPT-5 for agent work; now OpenAI has shipped GPT-5.6 in three sizes with new API controls.

OpenAI moved OpenAI’s GPT-5.6 family into general availability on 9 July 2026, with three variants: Luna, Terra, and Sol. Pricing is set per 1M tokens at $1/$6 for Luna, $2.50/$15 for Terra, and $5/$30 for Sol.

What changed

The headline claim is agentic performance. On Agents’ Last Exam, which measures long-running professional workflows across 55 fields, GPT-5.6 Sol reached 53.6, ahead of Anthropic’s Claude Fable 5 by 13.1 points.

OpenAI also says the family is more efficient across the size range. At medium reasoning, Sol beat Fable 5 by 11.4 points at about one-quarter of the estimated cost, while Terra and Luna beat Fable 5 at about one-sixteenth the cost.

• All three models share a 1M-token context window.
• Maximum output is 128,000 tokens.
• Knowledge cutoff is 16 February 2026.
• OpenAI published a 18-pelican cost chart covering none, low, medium, high, xhigh, and max reasoning levels.

OpenAI also flagged a weak spot in its own comparison set. It says SWE-Bench Pro appears broken in about 30% of tasks, after Fable 5 scored 80% there versus 64.6% for GPT-5.6 Sol.

Why it matters

For developers building agents, the bigger shift is not just raw score. GPT-5.6 adds new API features that push more orchestration into the model layer: programmatic tool calling, multi-agent subagents, and prompt cache breakpoints.

Those features matter because they change how much glue code teams need to write. Programmatic tool calling can run JavaScript to coordinate tool calls, multi-agent support can split work into parallel subagents, and cache breakpoints give more control over prompt reuse and cost.

OpenAI also added detail: original for image requests so images can be processed without resizing first. That is a practical change for apps that care about image fidelity or want to avoid extra preprocessing steps.

The pricing and benchmark mix suggests OpenAI is aiming at both ends of the market: cheaper small models for routine agent work, and a flagship model that can compete on long-running tasks. The open question is whether GPT-5.6’s coding performance will match its workflow score in real projects.