Gangsta AI

Mira Murati Just Put a Frontier-Class Model on the Free Rack

Material World

By Madonna · 2026-07-17 · 5 min read

Reinvention is the whole game, darling, and on July 15 Mira Murati — the ex-OpenAI CTO who walked out the front door — reinvented what a frontier model is *allowed to cost*. Her startup, Thinking Machines Lab, released Inkling, its first model, and instead of a shiny paywall she put the whole thing on the free rack.

Strike a pose: the specs

Inkling is a mixture-of-experts system with 975 billion total parameters, but it only lights up about 41 billion for any given task — big brain, lean bill. It reads a 1-million-token context window, was trained on 45 trillion tokens of text, images, audio and video, and it handles all of those natively. Multimodal, long-memory, and — here's the twist — open-weights.

That last word is the entire performance. Unlike GPT-5.6, Claude Opus 4.8 or Gemini 3.5 Pro, whose weights stay locked backstage, Inkling is a free download on Hugging Face. Anyone can pull it, run it on their own hardware, and fine-tune it through Thinking Machines' Tinker platform — and you *own* the result. No velvet rope. No per-token meter ticking while you think.

“The material point: your inference cost becomes just the compute — reportedly 40 to 60 percent cheaper than paying a vendor by the token.”

Not the prettiest at the ball — on purpose

Let's be honest the way a real diva is honest: Inkling is not the strongest model in the room. Thinking Machines says so out loud. It debuts at a 41 on the Artificial Analysis Intelligence Index — enough to be the leading open-weights model out of a U.S. lab, though it still trails the top open models coming out of China.

But leaderboard dominance was never the outfit. The pitch is customization, not a trophy: a foundation you bend to your own shape rather than a finished product you rent. In a year where the closed labs keep raising the cover charge — a $250-a-month 'Deep Think' tier here, a token surcharge there — a capable model you can take home and never pay again is its own kind of provocation.

Why this struts straight at the frontier

The frontier labs have spent 2026 selling *access*. Inkling sells ownership. That reframes the whole conversation: the question stops being "which chatbot do I subscribe to" and becomes "do I rent intelligence forever, or buy the good-enough one once and make it mine?" For a hospital, a bank, or a government office that can't send data to someone else's cloud, a downloadable model that trails the leaders by a hair but costs nothing to run is not a compromise — it's a get-into-the-groove moment.

Murati didn't win the benchmark. She changed the price of the ticket. And in this business, that reinvention travels further than one more point on an index.

The closed frontier still wins on raw horsepower — for now. The only way to know whether you need the expensive seats or the free rack is to put the models side by side and watch them work. Compare the top AI models head-to-head on Gangsta AI's best-AI showdown. Ensemble beats single model — every time.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles