On 22 September, Anthropic released Claude Opus 5.5. Minutes later, OpenAI came out with two new GPT-6 models, Sol and Luna, at half the price of their predecessors. Two announcements within an hour, both with a price cut as the main point. What does that mean for your own AI bill?
What has been announced exactly
According to SiliconANGLE, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. If you want faster processing, you pay double ($8 and $40). Cache reads drop to 20 cents, 60% lower than with Opus 5. Anthropic itself calculates that a typical workload comes out about 40% cheaper, partly because, by its own account, output is generated more than 30% faster. The model runs on the Claude Developer Platform, AWS, Google Cloud and Microsoft Azure.
OpenAI's Sol comes in at $2 per million input tokens and $10 per million output tokens. Luna sits an order of magnitude below that: 10 cents input, 50 cents output. Both are a halving compared with the GPT-5.6 versions with the same names; cache reads get a 90% discount. Luna is also available in the desktop app for free and Go users.
How to read this
Pay attention to who supplies the figures. The benchmark scores come from the providers themselves: Anthropic claims 58.7% on Terminal-Bench-Science 0.1 against 29% for Opus 5, while OpenAI reports 68.8% (Sol) and 66.6% (Luna) on DeepSWE v1.1. These are not independent measurements. Run your own test set with your own documents before you overhaul your entire software development or customer service.
That prices are falling is no coincidence, nor is it charity. Ara Kharazian, chief economist at Ramp, told Fortune: "OpenAI and Anthropic are fighting a price war that pushes the price of AI down and thereby their own ability to make a profit." Translated to your situation: don't assume the current rate will be the final rate. We already saw this year that introductory prices received an end date. Record in your budget which rate you are calculating with, and when you will look again.
And watch out for the trap we described last week with Grok 4.7: a lower rate per token is not a lower bill if the model thinks for longer and uses more tokens. Measure costs per completed task, not per million tokens.
What does NOT change
The GDPR picture remains as it was. These announcements concern price, speed and benchmarks, not data processing, retention periods or data residency. The terms that determine whether you may send healthcare, legal or government data through such a model are separate from this release. For the current profile per provider, see our model comparison and the safety page.
In short: cheaper calculation is a welcome bonus, but it does not change your data processing agreement or your DPIA.