OpenAI, the developer of ChatGPT, said on the 9th (local time) that it officially released its latest artificial intelligence (AI) model, "GPT-5.6."
OpenAI, at the request of the U.S. government, pre-released the GPT-5.6 lineup to a limited number of agencies over the past two weeks and made it available to the public on this day. GPT-5.6 consists of the high-performance top-tier model Sol, the second-tier model Terra, and Luna, which has strong expense efficiency.
OpenAI said the new model not only set top performance in areas such as coding, knowledge work, and scientific research, but it can perform the same tasks with fewer tokens and lower expense compared with competitors' top-tier models, improving "performance per dollar."
Sam Altman, OpenAI's chief executive officer (CEO), said in an interview with CNBC on this day, "GPT-5.6 Sol improved token efficiency for agentic coding tasks by 54%." A token is the basic unit an AI model uses to process information and generate responses. Altman noted, "Corporations care about what value they are getting in return for their AI expenditure," emphasizing that GPT-5.6 offers excellent cost efficiency relative to performance.
OpenAI rated Sol, the top-tier model, as "the best coding model ever." It also said its cybersecurity capabilities rival Anthropic's "Claude Mythos 5."
On Terminal-Bench 2.1, which measures coding ability in terminal environments, Sol scored 88.8%, edging out Mythos 5's 88% by a narrow margin, and in "CyberGym," which evaluates cybersecurity capability, it scored 84.5%, surpassing Mythos 5's 83.8%. On the "Agent's Last Exam" (ALE) metric, which measures practical agent abilities, it posted 52.7%, far ahead of Fable 5 (40.5%) and Opus 4.8 (45.2%).
However, on SWE-Bench Pro, which gauges programming language coding ability, it scored 64.6%, falling short of Mythos 5 (80.3%) and Opus 4.8 (69.2%).
Artificial Analysis, an AI model evaluation firm, gave GPT-5.6 Sol a score of 59, ranking it second after Fable 5 (60). GPT-5.6 Terra received 55 points, placing it fourth after Opus 4.8 (56).
Accordingly, Grok 4.5 from SpaceXAI, which debuted the previous day at No. 4, fell two spots to No. 6 in just one day.
The API (application programming interface) price for GPT-5.6, based on input and output values per 1 million tokens, is $5 and $30 for Sol, $2.5 and $15 for Terra, and $1 and $6 for Luna.