Frontier AI you can own, not just rent

For the past few years, building with the best AI meant renting it. You send data to OpenAI or Anthropic, pay per query, and live by their rules. This week, Thinking Machines Lab — founded by Mira Murati, the former chief technology officer of OpenAI — released Inkling: a frontier-class AI model you can download and run yourself. It is open-weight, meaning anyone can get the raw model files, use them, and modify them freely.

Inkling is enormous on paper — 975 billion parameters, the numbers that determine how an AI thinks — but a clever design means it only activates about 41 billion for any given task. That makes it far faster and cheaper to run than its total size suggests. In coding tests it matched the best available tools while using roughly one-third as many tokens — the small chunks of text an AI service counts to bill you — as competing models of equivalent power.

What makes this more than just another model launch is the strategy behind it. Thinking Machines is not racing to be the strongest model overall. It is betting that companies want AI they can tune for their own domain — their data, their language, their rules. An early test with investment firm Bridgewater Associates scored 84.7% on financial reasoning at roughly 14 times lower cost than a comparable closed model. That cost gap is enormous when you are processing millions of documents a month.

For anyone building software or automating work, the message is concrete. You can download Inkling from Hugging Face today, fine-tune it on your own data, and run it on your own servers — no subscription, no permission needed. Open AI's cost and control advantage is real now, not a future promise. The honest tradeoff: you take on the responsibility of keeping the model safe and well-behaved, which takes real engineering effort.

If your product relies on AI, this is the moment to run a real comparison. Take a workflow you currently send to a closed model, run it on Inkling, and measure quality and cost side by side. The performance gap has closed enough to make the test worth doing — and if the cost savings hold in your case, they could change what you are willing to pay.