OpenAI, Anthropic Pull Ahead in AI as Meta, SpaceXAI Push Lower-Cost Models

OpenAI launches high-performance GPT-5.6

Anthropic and OpenAI form AI's top tier

Meta and others roll out lower-cost models

OpenAI and Meta each released new artificial intelligence models on July 9, unveiling GPT-5.6 and Muse Spark 1.1, respectively. SpaceXAI, which released a new model a day earlier, is also competing on lower prices.

GPT-5.6 comes in three versions: top-tier Sol, mid-tier Terra and lower-cost Luna. Sol posted 88.8% accuracy on Terminal-Bench 2.1, a coding benchmark that measures performance in a developer terminal, edging past Anthropic's top model, Fable5, at 88%. On the Artificial Analysis Index, which combines nine benchmarks including Terminal-Bench 2.1, Sol scored 59, one point behind Fable5's top score of 60. In the Agent Final Exam, which puts AI agents through roughly 1,500 tests, Sol reached 52.7% accuracy, beating Fable5's 40.5%. It was also cheaper: the computing cost to complete the full test set was $1,087 for Sol, less than half of Fable5's $2,315.

Meta is betting on a much lower price, although its performance still trails Anthropic and OpenAI. Muse Spark 1.1, released three months after Meta unveiled its first AI model, Muse Spark, posted 69.2% accuracy on Terminal-Bench 2.1, far below Sol and Fable5. But its application programming interface pricing is far lower. Fable5 costs $10 for input and $50 for output per 1 million tokens, the basic unit of AI computing, while Muse Spark 1.1 costs just $1.25 to $4.25, about one-tenth as much.

SpaceXAI, renamed from xAI, released Grok 4.5 a day earlier. Its API pricing ranges from $2 to $6, underscoring its focus on cost efficiency. Grok 4.5 ranks sixth on the Artificial Analysis Index with a score of 54, behind five models from OpenAI and Anthropic.

As SpaceXAI, Meta and OpenAI unveiled new models in rapid succession, Google has not released a new one since May. Gemini 3.5 Flash, launched then, has fallen to 10th on the Artificial Analysis Index with a score of 50. Some in the industry say the AI model race is shifting from a three-way contest among Anthropic, OpenAI and Google. Anthropic and OpenAI now lead on performance, while Google, Meta and SpaceXAI chase them with large-scale infrastructure.

Kim In-yeop, Silicon Valley correspondent inside@hankyung.com