OpenAI's New Model Just Took Coding Back From Anthropic

OpenAI has hit back where it hurt most. It launched its GPT-5.6 family, and the new flagship, called Sol, edges out Anthropic's Claude Fable 5 on the leading coding-agent benchmark, 80 to 77.2, while using less than half as many tokens to do it. Coding was Anthropic's stronghold. Not this week.

The efficiency number matters as much as the score. Using half the tokens to reach a better result means the model is cheaper to run for the same work, which is exactly what businesses care about when they deploy AI at scale. A model that is both better and cheaper is a genuinely strong combination. Winning on cost is winning twice.

The family goes wider than one model. Alongside Sol, OpenAI released models called Terra and Luna for different jobs, plus a new generation of live voice models with big advances in how naturally AI can listen and speak. Sol itself is tuned for agentic work, where the AI operates a computer and completes tasks rather than just answering. The line-up covers coding, general use and voice.

The timing is pointed. Anthropic has been the darling of developers, has just overtaken OpenAI on revenue thanks largely to coding and enterprise work, and only recently got its top model back online after a government shutdown. OpenAI reclaiming the coding crown, however narrowly, is a direct strike at the business that made its rival strong. This was aimed squarely at Claude.

The honest caveat is that benchmark leads are fragile. The gap here is small, rivals ship new models every few weeks, and lab-reported scores often flatten out in real-world use where reliability and integration matter more. Developers also switch tools slowly once they are comfortable. A benchmark win is a headline, not a moat.

So OpenAI answered its rival's rise with a faster, cheaper, better coding model and a broader family behind it. Sol on top of the coding index, half the token cost, voice models alongside. The lead will not last long, it never does. But for now OpenAI has taken back the one benchmark Anthropic could call its own.