source: simon willison: the new gpt-5.6 family: luna, terra, sol
level: technical
openai released the gpt-5.6 family with three models: luna, terra, and sol, from smallest to largest. pricing per million tokens is $1/$6 for luna, $2.50/$15 for terra, and $5/$30 for sol. these prices compare to claude opus at $5/$25 and claude fable 5 at $10/$50, but direct cost comparisons are tricky because reasoning token counts vary widely across models for the same task.
on the agents' last exam benchmark, which tests long-running professional workflows across 55 fields, gpt-5.6 sol scored 53.6, beating claude fable 5 by 13.1 points. even at medium reasoning, it outperformed fable 5 by 11.4 points at roughly one-quarter the cost. the smaller models, terra and luna, also beat fable 5 at about one-sixteenth the cost. however, on swe-bench pro, fable 5 scored 80% while gpt-5.6 sol reached only 64.6%. openai had previously flagged issues with that benchmark, estimating 30% of its tasks are broken.
new api features include programmatic tool calling, which lets models compose and run javascript to orchestrate tool calls, and multi-agent support for spinning up parallel subagents. prompt cache breakpoints now allow explicit cache control, and image requests can use detail: original to avoid resizing. early hands-on testing by simon willison found gpt-5.6 sol competent but not clearly better than fable for complex coding tasks. the model guidance page shows reasoning effort levels from none to max across all three models, with costs ranging from 0.71 cents to 48.55 cents per request.
why it matters: the gpt-5.6 family offers strong agentic performance at lower costs, making advanced ai more accessible for long-running tasks, but its mixed coding benchmark results highlight the need for careful model selection.
source: simon willison: the new gpt-5.6 family: luna, terra, sol