The September 22nd clips put two models through Higgsfield's game-building workflow, but offer no controlled performance comparison.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [Higgsfield AI on X](https://x.com/higgsfield_ai/status/2102471046356177001)
Why it matters #
Higgsfield is using rival AI models to market the part it controls: the tools that generate assets and assemble playable games. The demos show that pitch, not a measured model winner.
Alex Mashrabov (@alexmashrabov) put Higgsfield's game-building tools between two competing AI models on September 22nd, when Higgsfield AI (@higgsfield_ai) posted samurai-game demos labeled Claude Opus 5.5 and GPT-6 Astra. The clips give Higgsfield a timely way to show what its visual-generation tools can contribute when different models handle the game-building work. They do not establish which model builds a better game.
https://x.com/higgsfield_ai/status/2102471046356177001 Mashrabov previously co-founded AI Factory, which Snap acquired, and went on to lead generative AI work at Snap. He founded Higgsfield with Yerzat Dulat (@codentropy), Higgsfield's chief technology officer. Their original business centered on AI-generated images and video. Games extend that work into something a user can play, giving Higgsfield another use for the assets it already generates.
A comparison inside Higgsfield's workflow
The distinction matters because Higgsfield supplies part of the result in either case. In its Games 2.0 guide, published September 7th, Higgsfield describes GPT-6 Astra as handling code, logic and orchestration while Higgsfield generates creative assets and provides the game-building workflow. Its earlier account of multiplayer game creation describes a similar division with a Claude model: the model writes and directs the game, while Higgsfield generates assets and hosts the result. Those descriptions explain why a game made with either model is also a demonstration of Higgsfield's own tools.
Higgsfield's September 22nd post calls the clips a comparison in 3D gaming. It does not give readers a common prompt, build logs, generation times, credit costs or playable links with which to compare the two samurai games. A video can show the appearance of a result; it cannot, by itself, test whether combat works reliably, how the game performs during play or how much intervention went into each build. There is no sound basis here for declaring a winner. ()
The timing is useful to Higgsfield. OpenAI introduced GPT-6 Astra earlier in September, and Higgsfield had already documented how to use it in Games 2.0. Claude Opus 5.5 became the other name in Higgsfield's September 22nd demonstration. Pairing the models in one post directs attention to the layer Higgsfield controls: the generation and assembly of characters, settings and other game assets, regardless of which model a creator chooses for reasoning and code. That is Higgsfield's product argument, rather than a finding that the two models perform equally.
Higgsfield's account of how it built Supercomputer makes the product strategy unusually clear. Higgsfield says its creative staff had been working with Claude, ChatGPT or Gemini on one screen and Higgsfield on another. Supercomputer was built to bring planning, code and visual generation into a coordinated workspace. The company says requests from users prompted it to add game creation. A model comparison made inside that workspace doubles as a pitch for keeping the whole job there.
The business stakes are larger than one pair of demos. In August, Higgsfield announced a $400 million Series B led by DST Global, with participants including Tribe Capital, Goldman Sachs Alternatives and Intel Capital. Higgsfield put its valuation at $5.4 billion and said it had reached $700 million in annualized revenue. That last figure is a run rate reported by Higgsfield, not revenue booked over a full year. The financing announcement also identified research, infrastructure and hiring as uses of the money.
For Mashrabov and Dulat, the samurai clips show where some of that product expansion is headed. Higgsfield already lets users specify a game's genre, visual style, setting and multiplayer option in Games 2.0. The test beyond a social video is whether people can repeatedly turn those controls and generated assets into games they want to play, at a cost they are willing to pay. September 22nd's post demonstrates the pitch; it does not answer that test.