Two models, not just two names
OpenAI’s GPT-6 API lineup puts two names on the menu: Sol for complex coding and agentic workflows, and Luna for focused, high-volume tasks. That is the company’s positioning, not a published rule that one model is “smart” and the other is “small.” The docs show both accept text and images, return text, support the same reasoning-effort settings, and advertise a 1,050,000-token context window with up to 128,000 output tokens.
The model cards also list different knowledge-cutoff dates—April 20, 2026 for Sol and May 18, 2026 for Luna. That is a detail worth checking for research-heavy tasks; it is not a benchmark of either model’s accuracy.
Luna is dramatically cheaper on tokens
At standard API rates, Sol is listed at $2 per million input tokens and $10 per million output tokens. Luna is $0.10 and $0.50 respectively: a twenty-fold price difference on each token category. Cached input and cache writes keep the same ratio. If a workflow mostly runs many short, well-scoped tasks, Luna’s listed rates could change the economics substantially.
But long context has a pricing catch. OpenAI says prompts above 272,000 input tokens are charged at twice the input and cache rates and 1.5 times the output rate for the full request. Both models advertise a million-token context window, but that does not mean every million-token run costs at the headline rate. Tool-specific charges may also apply, so token prices are not always the whole bill.
The docs don’t show the quality trade-off
Both model pages list the same reasoning-effort choices, from none through max, with medium as the default. OpenAI directs users to the Responses API for built-in tools and function calling; Chat Completions supports function calling only when reasoning effort is set to none.
What the public model pages do not provide is an apples-to-apples Sol-versus-Luna quality, latency, or throughput evaluation. “Most efficient” is OpenAI’s description of Luna, not independent evidence that it will meet a particular team’s quality bar. The model cards alone cannot show how often a cheaper run needs retries, extra tools, or human correction.
A small test beats a naming guess
Give both models the same representative tasks and compare accepted-answer quality, total tokens, tool charges, wall-clock time, retries, and intervention rate. Match reasoning effort and endpoint settings; include the long-context price threshold if your prompts approach it. Choose by total cost per acceptable result, not by family name or token rate alone.