The release of GPT-6 Sol and Luna comes as OpenAI faces intense competition on price and performance from a range of open-weight models developed in China, as well as from Anthropic.

Anthropic also this week released an updated version of Claude Opus, version 5.5, that cuts costs compared to the previous release two months ago.

Cutting costs

OpenAI said Sol and Luna make the capabilities of GPT-6 “more efficient and accessible”.

API access to the 6-series versions of Sol and Luna is being made available at half the cost of the previous 5.6 series, which OpenAI said was due to improvements in caching and inference.

It said the new models are more factually accurate compared to their predecessors, with a lower error rate in programming, an evaluation it said was based on analysing real-world conversations in which users flagged wrong answers.

Factual accuracy

A study released this month found that popular chatbots, including ChatGPT, gave wrong answers on financial questions an average of 56 percent of the time, rising to 88 percent for complex queries.

OpenAI and Anthropic are both preparing for IPOs to fuel their capital-hungry businesses, with Anthropic’s expected as soon as this autumn, while OpenAI is believed to be preparing a flotation for early next year.