Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war

Yesterday it was Grok 4.7 and MiMo v2.6. Today the spotlight moved again: Anthropic launched Claude Opus 5.5, then about an hour later OpenAI answered with GPT-6 Sol and GPT-6 Luna. The result is less a neat product launch cycle than a knife fight over token prices.

OpenAI’s new Sol and Luna models are roughly half the price of their GPT-5.6 equivalents. Luna is the eye-catcher: at $0.10 per million input tokens, $0.01 for cached input, and $0.50 for output, it lands far below the older GPT-5.6 Luna pricing. Sol got a similar cut, and GPT-5.6 also has a scheduled 25% price increase coming in November, which makes the gap even uglier for the older line.

Anthropic did not sit still. Claude Opus 5.5 drops from the old Opus pricing of $5/$25 per million tokens to $4/$20, and cached reads fall 60%. That matters in long agent conversations, where most of the input can come in at cached-token rates. But the bigger surprise is that OpenAI’s new Sol now matches that same $4/$20 level, while GPT-6 Luna undercuts almost everything else in the table.

The practical effect is clear: one of the few reasons to reach for GPT-5.6 Terra, at least on price, has evaporated. And the cheap end got brutally cheap. Willison says GPT-6 Luna is now among the least expensive models OpenAI has ever released, beaten only by GPT-4.1 Nano and GPT-5 Nano.

Not everything about Opus 5.5 looks tidy. In Willison’s “pelican riding a bicycle” test, the model on max thinking mode kept reasoning until it hit the 128,000-token output limit and failed to return anything at all. He tried twice, both times burning $2.56 and nearly 20 minutes. That is not exactly the sort of behavior that inspires confidence in a “max” setting.