OpenAI releases GPT-5.6 family with Luna, Terra, and Sol variants

OpenAI’s GPT-5.6 family launched with three variants—Luna, Terra, and Sol—positioned from smallest to largest. The models are priced per 1M input/output tokens at Luna $1/$6, Terra $2.50/$15, and Sol $5/$30, positioning them below the pricing of Anthropic’s Claude Fable 5 ($10/$50) and competitive with the Claude Opus series.

All three models share a February 16, 2026 knowledge cutoff, a 1M token context window, and a maximum output token limit of 128,000. OpenAI highlights long-running agentic performance, citing results on the Agents’ Last Exam benchmark, where GPT-5.6 Sol achieves a score of 53.6, surpassing Claude Fable 5 (adaptive reasoning) by 13.1 points. OpenAI also claims efficiency gains, noting that even at medium reasoning levels, GPT-5.6 Sol beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost, while the smaller variants, Terra and Luna, outperform Fable 5 at approximately one-sixteenth the cost.

OpenAI’s benchmark narrative is nuanced by a self-reported result on SWE-Bench Pro, where Claude Fable 5 scored 80% compared to GPT-5.6 Sol’s 64.6%. In response, OpenAI published a critique of SWE-Bench Pro, estimating that approximately 30% of its tasks are broken and advising developers to audit results carefully.

The GPT-5.6 release also introduces new API features, including programmatic tool calling, which allows models to compose and run JavaScript to orchestrate tool calls; multi-agent support for spinning up subagents to perform parallel, focused work; prompt cache breakpoints for explicit cache control; and an 'original' image detail level to avoid automatic resizing before processing.