AI models now come in tiers. The cheap ones are surprisingly good.

A few months ago, choosing an AI model meant picking the best one you could afford. That is changing fast. On July 9, OpenAI launched GPT-5.6 not as a single model but as three — Sol, Terra, and Luna — each built for a different kind of task and priced to match. The family spans a five-times price range: from $1 per million tokens (the chunks of text an AI reads and writes) for Luna, up to $5 for Sol. And the most important part: the cheaper tiers already beat last year's frontier.

Sol handles the hardest work: complex reasoning, security analysis, autonomous agents that run for long periods. Terra covers solid everyday production tasks at half Sol's price. Luna does fast, repetitive jobs — classifying data, writing short replies, routing requests. OpenAI claims Terra and Luna already outperform Claude Fable 5, Anthropic's previous top model, at one-sixteenth the price with similar results. That claim is contested on some benchmarks — Claude Fable 5 still leads on software engineering tasks — but the direction is clear: capable AI keeps getting cheaper.

OpenAI is not alone. Anthropic launched Claude Sonnet 5 at $2 per million input tokens — cheaper than Sol — and describes it as near-flagship quality. Google and xAI are building the same way. Across the board, the market has shifted from one model per lab to whole families with very different prices — and the cost of good AI has fallen fast.

For anyone building products or running automated workflows, the practical move is to think in tiers. Most AI calls in a real product are not hard reasoning problems — they are simple lookups, summaries, or classifications. Routing those to Luna or an equivalent cheap model, and saving Sol-class models for genuinely difficult tasks, can cut your AI costs by 80% or more without losing quality where it counts. The question is no longer which company has the best AI. It is whether your team is routing each job to the right model.