Opus 5.5 and GPT-6 Sol: the model war moves to the price list

The new models from Anthropic and OpenAI bring no breakthrough in capability. They bring price cuts, and right now that is the main front of the competition. Opus 5.5 is the latest version of Anthropic's main model for coding and complex knowledge work. It costs 4 dollars per million input tokens and 20 dollars per million output tokens, 20 percent less than Opus 5. Cache reads now cost 0.20 dollars per million tokens, down 60 percent. In agent work and coding, cache reads make up most of the bill. The model also answers more than 30 percent faster.

Anthropic says the saving reaches 40 percent on typical tasks and default settings, because the model burns fewer tokens to finish the same job. One caveat applies to what the company calls high-risk areas such as cybersecurity and biology. Queries from those fields may be routed automatically to the older model.

OpenAI answers with GPT-6 Sol and Luna, the lighter variants in the GPT-6 family. Sol is meant to be an everyday tool for tasks that need decent quality at a lower cost. Luna is cheaper and smaller still. The backdrop is that the companies are racing not only against each other but against open models. Businesses increasingly reach for model routers so they can lean on expensive frontier models less often.

In China the same mechanism shows up in DeepSeek's price list. From 10 September, from 12:00 Beijing time, the flash model is served at new rates. Input with a cache hit fell in off-peak hours from 0.05 to 0.02 yuan per million tokens, a drop of 60 percent. Input without a hit fell from 1.5 to 1 yuan, one third less. Output fell from 4.5 to 4 yuan, about 11 percent less. Peak-hour rates are twice the off-peak ones.

The cut is telling because on 17 August DeepSeek had introduced peak tariffs and raised the output price from 2 yuan to 4.5 yuan in off-peak hours and 9 yuan at peak, an increase of 125 percent and 350 percent. A little over a month later the company worked out that it could cut prices without losing revenue. On 19 September it added that weekends and public holidays in China are billed around the clock at off-peak rates, which for many teams means a lower bill still.

For customers, though, the level of the rates matters less than the direction they are moving in. A price cut on top-tier models changes the maths not only in the labs but for everyone building their own products on them. What a year ago was an experiment now fits a small team's budget. The price war does not mean models get cheaper forever, since suppliers shift the cost onto weaker versions and onto peak-hour terms. Anyone who can move the load into the off-peak window pays less today than anyone paid a year ago for a comparable volume.