30초 요약
On the outdoor stage at OpenAI DevDay 2026, an unnamed host interviews Alex Mashrabov, co-founder and CEO of Higgsfield. Mashrabov describes Higgsfield as a multimedia AI platform for end-to-end advertising campaigns, plus an API so developers can ship interactive shows that used to take large production teams. The hard problem, he says, is not a single concept image but a continuous experience, which is why they built Supercomputer and an agent that chooses models, self-evaluates quality, and launches its own GPU infrastructure. He argues that OpenAI's Astra is the generation of models that unlocks computer use, collapsing a 15-person, 10-to-12-tool production tax so three-to-five-person teams can hit a gaming-level visual bar. Proof point: a live, 10-hour AI-generated stream of two creators walking New York, with Astra reconstructing the world. Higgsfield is exposing the same perimeter-GPU, guardrailed setup through its API, has open-sourced an AI video-editing skill it uses internally, and closes with a simple pitch: they built the company in 18 months, and now is the best time to build yours.
- 1A platform, then an APIHiggsfield started as a place for marketers to build ad campaigns end-to-end with Astra and Higgsfield's own image and video models. The API is the next bet: developers shipping interactive media that a single person could not have made a few years ago.
- 2Supercomputer, because one image is not a showGetting a concept image is easy. Delivering a continuous experience around it is not. Mashrabov says that gap is why they shipped Supercomputer and had to teach an agent to launch its own GPU infrastructure.
- 3$30,000 to make one skill production-readyA video-editing skill that can produce a consistent 10-minute video was impossible without Astra and cost more than $30,000 to deliver. Putting that into production means deciding which model to use, and agents that can score their own output.
- 4Astra unlocks computer useTraditional creative work needed 15-plus people fluent in 10 or 12 tools. Mashrabov calls Astra the next generation of models that can run those programs on a creator's behalf, which is how smaller teams reach a gaming-level visual bar.
- 5A live 10-hour New York streamHiggsfield spokesperson Adil and streamer Neon walked through New York, interacting with people, while Astra reconstructed the world in real time. The stream ran more than 10 hours. Mashrabov says two or three developers can now ship that form of media.
- 6Guardrails and perimeter GPUsAt larger scale, controllability matters. The agent sits inside guardrails and deploys through perimeter GPU infrastructure, which sped up Higgsfield's own enterprise go-to-market and is now in the public API.
- 7An open-sourced explainer skillThey open-sourced an AI video-editing skill for YouTube explainers, then used it internally to teach Higgsfield. Mashrabov jokes it is a kind of singularity: Astra helped them build the skill they use to make more videos about Higgsfield, and other developers now use it too.
핵심 정리
- Higgsfield is a multimedia AI platform for end-to-end ad campaigns, with an API so a one-person team can ship interactive shows using Higgsfield models and OpenAI Astra.
- A good concept image is easy; a continuous experience is not. Supercomputer exists so an agent can choose models, define quality criteria, and launch its own GPUs.
- One video-editing skill that produces a consistent 10-minute video cost more than $30,000 to deliver, so production agents have to self-evaluate which output is acceptable.
- Astra, in Mashrabov's view, is the generation of models that unlocks computer use, calling 10 or 12 creative tools on a creator's behalf and cutting a 15-person production tax.
- They ran a live 10-hour AI-generated stream of two creators walking New York City, with Astra reconstructing the world so it felt realistic.
- Enterprise use needs controllability: guardrails and secure perimeter GPU infrastructure, now available through the Higgsfield API.
- Higgsfield open-sourced an AI video-editing skill it uses to make its own product explainers; other developers now use the same skill for their tutorials.
챕터별 전체 원고 읽기
Welcome and introduction
The host opens the outdoor stage at OpenAI DevDay 2026 and introduces Alex Mashrabov as Higgsfield's co-founder and CEO. He sketches Mashrabov's path from Snapchat Cameos and generative AI at Snap to agents that choose models and spin up GPUs.
The host welcomes everyone to the outdoor stage at OpenAI DevDay 2026. He is excited for the next guest: Alex Mashrabov, co-founder and CEO of Higgsfield, who co-founded a product that became Snapchat Cameos and later led generative AI at Snap. What Mashrabov is doing now, the host says, is using agents to choose models and spin up new GPUs, unlocking new ways to create and experience media. He asks the crowd for a warm welcome.
The host says he is a huge fan but is not sure everybody knows the company, and asks how Mashrabov describes Higgsfield these days and what people are building with it. Mashrabov calls it a leading multimedia AI platform where creative marketers build advertising campaigns end-to-end. They use modern models like Astra and advanced image and video models from Higgsfield. Most importantly, those AI-generated ads become a way businesses actually grow revenue and sales.
What Higgsfield is for
Mashrabov describes Higgsfield as a multimedia AI platform where marketers build campaigns end-to-end with Astra and Higgsfield's own models, aimed at growing revenue. Unmet demand for interactive media led to a developer API so a one-person team can ship an interactive show.
He also felt there were unmet needs: a lot of people want to build new forms of creative and interactive media that were not possible before. That is why Higgsfield launched an API, so developers can ship those experiences and build new businesses. Combined with Higgsfield's models and OpenAI Astra, a team of just one person can ship an interactive show that would have been impossible to imagine a few years ago.
Why they built an agent
The host asks what forced the agent. Mashrabov says a concept image is easy and a continuous experience is hard, which is why they shipped Supercomputer. A 10-minute video-editing skill depended on Astra, cost more than $30,000, and needs agents that self-evaluate and pick models as they go.
The host notes that they went and built an agent, and asks what forced them to think that through. Mashrabov says they saw the same pattern over and over: it is very easy to get a good asset, like a concept image, and very difficult to deliver a continuous experience around it. That is why they shipped Supercomputer. On another podcast he had described a new skill — video editing — that can go and make a consistent 10-minute video, something that would be impossible to imagine without Astra.
They spent over $30,000 to deliver that experience. To put skills like that into production, a lot of decisions have to be made about which model to use for which use case. That is also where agentic systems have to be really good at self-evaluation: defining specific criteria for when the quality of the output is acceptable, then choosing those models as they go.
From spark to GPU infrastructure
Asked how a creative spark becomes a finished product, Mashrabov points to creativity in the Sol model and Astra, then to a specific execution plan and quality criteria. Deterministic workflows and fine-tuned video models matter too, so Supercomputer's agent had to launch its own GPUs.
The host asks him to go high level: you have a creative spark, and then an end product comes out — what happens in between? Mashrabov says that is a good question, because one of the areas where the Sol model and Astra are truly unique, starting probably from Sol 5.5, is creativity. It all starts with an idea and a concept, but then it is important to bring that concept into a very specific plan of execution and define very specific criteria for what the final outcome is.
They have also seen that deterministic workflows are valuable. Those workflows may include advanced rendering with video models, including fine-tuned ones. That is why, on Supercomputer, they had to build an agent that can launch its own GPU infrastructure. Mashrabov says that has been critically important for delivering consistent execution across use cases.
Astra and the production tax
Mashrabov argues Astra is the generation of models that unlocks computer use. Creative production used to need 15-plus people across 10 or 12 tools. In gaming especially, visual quality and budgets have soared; Astra can call those programs so smaller teams can compete.
The host asks where else in the stack OpenAI models fit. Mashrabov says what he has seen with Astra is that it is probably the next generation of models that actually unlock computer-use cases. Traditionally, a creative workflow required a team of more than 15 people, each proficient at maybe 10 or 12 tools, which put a huge production tax between the idea and the execution.
He has seen that especially in gaming, where the bar for visual quality keeps getting higher and budgets have grown enormously. That is where he feels Astra can let smaller teams be much more successful, because Astra can call all those 10 or 12 programs on the creator's behalf. The host boils it down to sub-agent work.
Real-time video and a 10-hour stream
The main trade-off was speed: continuous generation and automatic editing have to feel real-time. Working with streamer Neon, Higgsfield ran a live 10-hour AI-generated walk through New York with spokesperson Adil, Astra reconstructing the world, a format two or three developers can now ship.
Asked about the bigger decisions and trade-offs, Mashrabov goes back to those examples: unlocking continuous video generation and automatic video editing. That takes many decisions, and they really had to optimize for speed in order to enable real-time experiences. Ultra-fast models announced the same day are making that more accessible to everyone. Back then, they were working with Neon, a popular streamer, to unlock the first AI-generated stream.
The host asks him to tell that story. Mashrabov says it is truly unique: two real creators — Adil, Higgsfield's spokesperson, and Neon — walking through New York City, interacting with other people, while Astra reconstructed the whole world so it felt realistic. Live, the host checks. Live, Mashrabov says, and the stream went over 10 hours. That is a new form of media that would have been impossible to imagine before.
And now, he says, every developer in the world — a team of two or three people — can ship these experiences and build businesses on top of them. The host likes that as a customer use case. With an audience full of developers, he asks for a specific moment that might inspire them, because going to 10 hours is amazing.
Small teams and enterprise control
Mashrabov has watched three-to-five-person teams ship direct-to-consumer apps on the Higgsfield API. At larger scale, the agent needs guardrails and secure perimeter GPU infrastructure. That setup accelerated Higgsfield's enterprise go-to-market and is now available to other builders.
What has been truly unique, Mashrabov says, is how many of these moments there are: teams of three or five people shipping a product end-to-end. He has seen a lot of that in direct-to-consumer apps, mobile apps and websites — entirely new consumer experiences enabled by the Higgsfield API and OpenAI models.
He also wants to step back toward larger-scale use cases, where controllability becomes important. It mattered that the agent does not just create, but has guardrails around it and can provide a secure deployment through perimeter infrastructure with GPUs. That helped Higgsfield accelerate its own enterprise go-to-market. Now that this is available in the Higgsfield API, others can build those experiences as well.
The host says this is what they always recommend, using a hockey metaphor: if you are building with the latest models, skate to where the puck is going. Higgsfield does that well — building the experience when the model is not quite there yet, going from two hours to ten, until suddenly it is possible and everybody can access it. For the developers in the room and watching at home, he asks what they should be building today to get ready for the next AI jump.
What to build next
The host asks what to build for the next jump. Mashrabov says interactive media creation is being democratized as models improve, and Higgsfield has open-sourced an AI video-editing skill it uses internally. He closes: they built Higgsfield in 18 months, and now is the best time to build yours.
Look, Mashrabov says, the models improve so quickly. Even earlier that day they had seen new capabilities around decisioning. What truly excites him is that interactive media creation is being democratized like never before. Historically it was bottlenecked by the ability to coordinate a lot of tools; with Astra, it finally becomes possible to unlock that.
What also excites him is that they open-sourced new skills of their own, including an AI video-editing skill to help make AI YouTube explainer videos. They use it at Higgsfield as well. The skill was valuable for building more education about Higgsfield products, and it is funny, he says, that Astra helped them build a skill they use internally to make more videos about how to use Higgsfield — a kind of singularity. What matters is that developers now use the same skill to build tutorials about their own products.
The host calls it a brand-new use case, says what they are building is incredible, and hopes everybody feels inspired to build a little bit. Any last words for the audience? In just 18 months, Mashrabov says, we built Higgsfield, and now is the best time to build yours. Thanks all around.
