OpenAI previewed GPT-5.6 in three sizes, Sol at the top, Terra in the middle, Luna for cheap volume. The catch is who gets in: roughly 20 organizations…
OpenAI previewed GPT-5.6 in three sizes, Sol at the top, Terra in the middle, Luna for cheap volume. The catch is who gets in: roughly 20 organizations…
OpenAI previewed GPT-5.6 in three sizes, Sol at the top, Terra in the middle, Luna for cheap volume. The catch is who gets in: roughly 20 organizations, each cleared by the US government before they can touch it. All three are rated high-risk for cyber and bio, so broad access waits weeks. For the first time the gap between a model existing and you being allowed to run it is a real delay shaped by regulators, not a signup form.
In one week DeepMind lost Noam Shazeer, a co-author of the Transformer paper, to OpenAI, and John Jumper, the Nobel laureate behind AlphaFold, to Anthropic. Two more AlphaFold and Gemini contributors are reportedly following Jumper. Alphabet shed about $250 billion in value on the news. Talent is concentrating at the two labs already winning, and the people who build frontier models are telling you where they think this goes.
Anthropic's run-rate revenue pushed past $30 billion, and the shape underneath matters more than the headline. Over 1,000 customers now spend more than $1 million a year on Claude, a count that doubled in under two months, and eight of the Fortune 10 are on board. This is not pilots anymore. Claude is a budgeted line in large companies, which makes it the safe default a procurement team can defend.
OpenAI unveiled Jalapeno, its first custom AI chip built with Broadcom, days after Anthropic locked multiple gigawatts of next-generation compute through Google and Broadcom. The pattern is clear: the frontier labs are buying their own supply chain instead of leasing it from whoever has GPUs. For anyone building on top, it means your provider's cost curve, and eventually your prices, get set by silicon they own rather than rent.