Following on from Claude Opus 5.5 comes Gpt 6 Sol e Luna. But weren’t the AI giants supposed to be slowing down?

Yesterday, Anthropic unveiled Claude Opus 5.5. A few hours later, OpenAI launched GPT-6 Sol and Luna. The heads of both companies are calling for regulations, safety standards and a cautious approach to the development of artificial intelligence. Then, no sooner had the sentence been finished than another model arrived. For now, the slowdown is more in words than in deeds. No one is slowing down, but almost everyone is raising their hand to ask for help – or to ask for time whilst waiting for a more solid foundation. Perhaps.

Having said that, to understand what changes with GPT-6, it’s best to start with the family. Astra, unveiled at the start of the month, is the flagship model: OpenAI recommends it for complex tasks, from writing code to using computers and business applications. Sol brings some of those capabilities to demanding but more frequent tasks. Luna is the lightweight version, designed to handle large volumes of tasks at a low cost. All three can work with tools; Astra remains the one to turn to when the problem requires more resources.

The easiest change to measure is the price. In the APIs, for one million tokens, Sol costs $2 for incoming requests and $10 for outgoing ones; Luna costs 10 and 50 cents. OpenAI refers to a 50 per cent reduction compared with the promotional prices of the corresponding GPT-5.6 models. This makes a real difference for those who send thousands of requests to the AI every day: if each response costs less, it becomes more feasible to carry out more attempts, checks and workflow steps.

The behaviour promised by OpenAI is also changing. Sol and Luna have inherited improvements from Astra in terms of computer usage, programming and following instructions. According to an internal company test based on conversations in which users had reported errors, Sol makes around half as many factual errors as its predecessor.

In terms of performance, GPT-6 Sol and Luna mark a clear shift towards computational efficiency rather than pure brute force: they inherit from the Astra architecture the ability to handle agent-based workflows, complex programming and direct interaction with operating systems (computer-use), but with significantly reduced latency, halved hallucinations compared to the 5.x generation, and reasoning synthesis that requires a fraction of the output tokens previously needed.