GPT-5.6: OpenAI promises more performance with less token consumption
OpenAI releases GPT-5.6. According to the manufacturer, its new AI models outperform the competition from Anthropic and consume fewer tokens.
(Image: Camilo Concha/Shutterstock.com)
Under strict US government regulations, OpenAI is releasing GPT-5.6, its new generation of AI models. According to OpenAI, it can match Anthropic's Mythos 5 in numerous benchmarks – and like Mythos, it was withdrawn for the general public under pressure from the US government. GPT-5.6 is said to exhibit improved agentic capabilities in programming, biology, and cybersecurity. The protective measures against misuse have also reportedly been improved.
GPT-5.6 includes the Balanced model Terra for everyday use, Luna, which is designed for speed and efficiency, and Sol, the most powerful model as the flagship. OpenAI cites results from various benchmark tests to underscore the strengths of GPT-5.6. In the coding benchmark Terminal-Bench 2.1, Sol and Sol Ultra outperform Mythos 5 with scores of 91.9 percent and 88.8 percent, respectively, compared to Mythos 5's 88 percent. Sol Ultra is an enhanced version of Sol with a new feature that allows it to call upon additional sub-agents to further increase Sol's performance.
In the biological domain, OpenAI names the benchmark GeneBench v1, comparing itself here with the previous model GPT 5.5. When using 20,000 tokens, Sol achieves around 25 percent, while GPT-5.5 only achieves 20 percent.
According to OpenAI, much lower token consumption than Mythos 5
Token consumption for cybersecurity tasks is also said to be significantly lower. In the ExploitGym benchmark, Sol, Terra, and Luna are said to achieve the same performance as Mythos Preview, using only about a third of the tokens. It is not clear which Anthropic model is hidden behind Mythos Preview.
OpenAI has also revised its protective measures against misuse, according to its own statements. GPT‑5.6 is supposed to refuse prohibited cyber support, even if users try to jailbreak the model, i.e., bypass the protective measures. Real-time classifiers are also supposed to monitor the responses while GPT-5.6 is generating them.
Videos by heise
In terms of pricing, GPT-5.6 models are said to undercut Anthropic's competition: 2.50 US dollars per million input tokens, and 30 dollars for output. For the same amount, Fable 5 and Mythos 5 both cost 10 dollars for input tokens and 50 dollars for output tokens.
(nen)
