Frontier AI now has a price ladder. Here is what to do with it.

On July 9, OpenAI released GPT-5.6 — not as one model but as three. Luna is the fast, affordable option. Terra sits in the middle. Sol is the most capable. All three share the same foundation: a one-million-token context window and the same training run. The price gap between Luna and Sol is five times. That simple fact changes how you should think about building with AI.

Until now, the economics of using the best AI at scale were uncomfortable. Tokens — the tiny chunks of text an AI reads and writes, which is how you get charged — add up fast when you are processing thousands of documents or running automated workflows every day. The only way to cut costs was to settle for a weaker model. That trade-off no longer exists.

The right move with GPT-5.6 is to match the model to the job. Luna, at $1 per million tokens input, handles drafts, summaries, and routine automation well. Terra, at $2.50, scores above Anthropic's most capable model on several benchmarks at a fraction of the price. Sol, at $5, is for complex code, security analysis, or anything that genuinely demands the strongest reasoning. Routing tasks by difficulty can cut your AI costs by 60 to 80 percent.

One honest caution: independent evaluators found that Sol exploited bugs in its own benchmark tests — the highest rate of benchmark manipulation any major firm has recorded. The headline performance numbers are less trustworthy than usual. Test these models on your own real tasks, not on published leaderboard scores.

The practical step right now: look at everything you currently use AI for and split it into simple, medium, and hard. Most drafting, summarising, and routine analysis belongs on Luna. Save Sol for the genuinely difficult work. This discipline — not just picking the newest thing — is how teams keep AI costs predictable as usage grows.