OpenAI renews its challenge to Anthropic: Sol, Terra and Luna – the new GPT-5.6 models for coding and cybersecurity – are here
-U60684414814Kct-1440x752@IlSole24Ore-Web.jpg?r=1170x507 "OpenAI renews its challenge to Anthropic: Sol, Terra and Luna – the new GPT-5.6 models for coding and cybersecurity – are here")
Publish Date: 2026-06-29 02:47:00
Source Domain: en.ilsole24ore.com
Using an unordered list, summarize the following article with between 4 and 8 key points. They are called Sol, Terra and Luna, and are the three versions of GPT-5.6 released in preview to ‘a select group of trusted partners and organisations’ by OpenAI. And, at least the first of these, reignites the rivalry with Anthropic and its flagship model, Claude Mythos – which is also not available to the general public.Luna is described as a balance between performance and cost; Terra is comparable to ChatGPT-5.5 but at half the cost; Sol is described in the press release announcing its launch as a ‘next-generation model’. This press release is nothing more than a benchmark comparison between OpenAI’s LLMs and those of Anthropic.Meanwhile, in terms of performance: Sol’s success rate in writing code, as assessed by Terminal-Bench – which is now one of the most widely used tools for comparing AI models – stands at 88.8 per cent and rises to 91.9 per cent when the new ‘Ultra’ reasoning mode is selected. In practice, this is a mode of operation in which an agent coordinates the activities of various sub-agents to achieve the ultimate goal. And what about Claude Mythos 5? It stops at 88 per cent.That’s not all: GPT-5.6 Terra matches the performance of Claude Fable 5, with a code-writing accuracy rate of 84.3 per cent. Meanwhile, Luna achieves 82.5 per cent, surpassing the 78.9 per cent achieved by Claude Opus 4.8, Anthropic’s most powerful model currently available.Furthermore, when it comes to cybersecurity, Sol achieves the same results as Mythos whilst using only a third of the tokens. This result was measured using ExploitGym, a benchmarking tool developed by OpenAI itself in collaboration with researchers at the University of Berkeley.
