Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output, roughly 20% below Opus 5, with cache reads at $0.20 per million, about 60% cheaper, per Simon Willison. About an hour later OpenAI released GPT-6 Sol at $2 and $10, exactly half of Opus 5.5 on both sides, and GPT-6 Luna at $0.10 and $0.50, per 9to5Google. Sol replaces GPT-5.6 promotional pricing of $4 and $20.
Sol at $2 and $10 is exactly half of Opus 5.5 on both sides. That is pricing set against a rival's rate card, published within an hour of Anthropic's launch, with no cost pass-through behind it. Anthropic's move was quieter and aimed at one workload: cutting cache reads 60%, to $0.20, pays off most in long agent sessions that re-read the same context. That is the workload both companies are chasing.
For buyers the comparison stays awkward, for the reason we covered on Tuesday. Grok 4.7 kept a $2 and $6 rate card while spending about 81,000 output tokens per task, which left it costlier per finished job than a model with a higher list price. A 50% list-price cut moves your bill only if reasoning length holds steady. Run your own workload first.
Bottom Line
Two labs cut flagship prices within an hour, and the cheap tier now starts at $0.10 per million input tokens. Re-price your workloads on completed tasks, because list prices have become a competitive weapon and say less about what you will actually pay.