The new AI vendor divide: finished model, or platform you reshape?

Frontier AI vendors are sorting into two tribes — in the framing this story applies to Thinking Machines' strategy — and the safety boundary moves with the choice. One camp sells a finished model and keeps the liability. The other sells a platform you reshape and hands the liability to you.

Thinking Machines Lab released Inkling, its first open-weight model, on July 15, 2026, and the lab put the trade-off in plain language. They are not selling the smartest model available. They are selling one buyers can fine-tune to their domain and deploy under their own controls, and they are saying so on the record.

When the next enterprise pitch lands, ask which tribe the vendor belongs to. If they sell you a finished model, they own the safety review, the red-team work, the regulatory exposure. If they sell you a platform you reshape, all of that lands on you, plus the ML talent bill. Thinking Machines' Tinker customization layer requires serious in-house machine-learning capability to run. The vendor's contribution is the starting point; the customer's contribution is everything that comes after.

The bet is that buyers will accept the trade. Give up best-in-class capability. Take on the reshape work. Own what you build. Murati is the first frontier lab founder to make the concession out loud. She will not be the last.