OpenAI has unveiled its new family of models - GPT-5.6, in three versions: Sol (the workhorse), Terra (the mid-tier option) and Luna (the cheapest). The company promises power in business work, coding and scientific research, and chief executive Sam Altman claims the new models are "orders of magnitude" more efficient than the previous ones - Sol reportedly using 54 percent fewer tokens when coding.

Put simply, this is another big model thrown into an already crowded field. This week SpaceXAI and Meta released similar announcements too, so GPT-5.6 does not enter an empty stage - it enters a scrum where everyone claims to be the best. And here comes the most interesting part: OpenAI's marketing is tailored above all to hit Anthropic, the rival that managed to cast itself as the likeable underdog in the race and, precisely for that reason, is quickly winning over business clients.

OpenAI holds nothing back in the comparisons. Citing a well-known benchmark - the Artificial Analysis Coding Agent Index - the company claims Sol is its "best coding model yet" and that it beats Anthropic's Fable 5 with a score of 80 versus 2.8 points lower, while using less than half the tokens and a third less money. Terra, they say, is slightly above Fable 5, and Luna beat Opus 4.8. Are these numbers worth as much as they sound? The benchmarks are chosen by whoever is bragging about them - an old rule that holds in tech too.

More interesting than the benchmarks is what sits behind the model. OpenAI calls GPT-5.6 its "strongest cybersecurity model yet" - and that is exactly what previously worried the Trump administration, which wanted to restrict the release over fears of abuse. The model supports defensive activities: threat modelling, code review and patching, simulated attacks on one's own systems. The same tool that protects can also attack - a dilemma as old as cybersecurity itself.

Along with the models, OpenAI also launched ChatGPT Work - a work companion for business teams on desktop, web and mobile, helping with everyday office tasks like drafting documents, spreadsheets and presentations. The prices are clear: per million tokens, Sol costs 5 dollars input and 30 output, Terra 2.50 and 15, and Luna 1 and 6. The numbers are precise, the promises are big, and the real test - as always - will not come from press releases, but from the people who work with these models every day.