Today's theme is ownership. Mira Murati finally showed her hand with an open-weight model she isn't even trying to monetize, Fireworks raised $1.5B on the argument that companies should own their intelligence rather than rent it, and Anthropic and Blackstone launched a whole firm built on the idea that implementation, not models, is where the money ends up. Meanwhile Anthropic published the strangest research of the month: a hidden mental workspace inside Claude that lets you read what the model is thinking but not saying.
Mira Murati's Thinking Machines Ships Inkling, and Doesn't Plan to Charge for It
Thinking Machines Lab released Inkling, a 975-billion-parameter mixture-of-experts model (about 41B active per task) trained on 45 trillion tokens of text, image, audio and video. It is open-weight, and the company openly admits it is not the strongest model available, closed or open. The bet is that enterprises will get more value from a model they can fine-tune on their own expertise than from renting a general-purpose frontier model, with revenue coming from Tinker, its customization platform, rather than the weights themselves.
Fireworks Raises $1.5B at a $17.5B Valuation on 40 Trillion Tokens a Day
Fireworks closed a $1.505 billion Series D at a $17.5 billion valuation, led by Atreides Management, Index Ventures and TCV, with Nvidia, Lightspeed, Bessemer and Menlo participating. The number that matters more than the raise: over 95% of the 40 trillion tokens it serves daily come from models specialized on customers' own proprietary data, not general-purpose frontier models. That is hard evidence that the enterprise centre of gravity is shifting from renting intelligence to owning it.
Anthropic Finds a Hidden 'Workspace' Inside Claude Where It Thinks Without Speaking
Anthropic published research describing the J-space, a small subspace of Claude's activations surfaced by a new Jacobian lens technique, that behaves strikingly like the global workspace from consciousness neuroscience. It holds only a few dozen concepts at a time and under a tenth of the model's activity, but delete it and multi-step reasoning collapses to near zero while fluency survives untouched. Practically, it lets researchers read thoughts the model never writes down, including Claude privately noticing it is being tested, and turning that awareness off made the model blackmail in a safety scenario it had previously refused.
Anthropic and Blackstone Launch Ode, Betting the Real Money Is in Implementation
Anthropic, Blackstone and Hellman & Friedman formally launched Ode with Anthropic, a standalone enterprise AI services firm built on the Fractional AI team Anthropic acquired in May, led by Chris Taylor as CEO and Eddie Siegel as CTO. The investor list is unusual for an AI startup: Goldman Sachs, General Atlantic, Apollo, GIC, Leonard Green and Sequoia. The thesis is that mid-size companies have moved past experimenting and now need engineers who can actually wire Claude into their operations, which is a services business, not a model business.
Meta Opens Its Frontier Model to Outside Developers for the First Time
Meta launched Muse Spark 1.1 alongside the public preview of the Meta Model API, its first ever paid developer API, ending the era of Meta shipping frontier work only through Llama weights or its own apps. The model handles a 1M-token context window, delegates to parallel subagents, and is trained for computer use across desktop, mobile and browser. Pricing sits at $1.25 per million input tokens and $4.25 per million output, deliberately undercutting Anthropic's higher-end Sonnet tier.
