The demo that got engineers talking wasn’t a chatbot answering a question. It was a live camera feed. OpenAI this week launched GPT-5, its latest flagship model, and the headline feature is one the company has not shipped before: real-time understanding of streaming video. The model watches a live picture and interprets what is happening as it happens, rather than analyzing footage after the fact.
OpenAI introduced the model across ChatGPT and its API in a phased rollout, positioning GPT-5 as a unified system that routes simple questions to a fast model and harder ones to a deeper reasoning model. For developers, the company released three API sizes, gpt-5, gpt-5-mini and gpt-5-nano, priced from $1.25 per million input tokens down to $0.05, with output pricing from $10 to $0.40 per million tokens.
The new video capability is the piece that stands apart. TechCrunch, reviewing the launch, called it a defining step in multimodal AI’s move from static images and text to dynamic, real-time interaction. In practice, the model ingests a live stream and returns interpretations within seconds — a loop from perception to reasoning that OpenAI says has moved close to practical usefulness. On internal video benchmarks, the company reports GPT-5 scoring 84.6% on VideoMMMU, a widely used video-reasoning test, ahead of where earlier models stood on the same task.
What real-time video changes
The distinction is technical as well as practical. Video is temporal — meaning unfolds in time. Earlier multimodal models treated video as a collection of still frames, extracting a picture every few seconds and analyzing it in isolation. GPT-5, according to OpenAI’s technical materials, reasons over the stream itself, which lets it understand motion, sequence and cause and effect. That difference matters in any application where timing is the point: a hand reaching toward a switch, a car drifting across a lane, a weld forming on a production line.
The shift for machine vision is from inspection to monitoring. An inspection system photographs each part and checks it. A monitoring system watches continuously and notices when something changes. The second mode is what live video understanding enables, and it is why industrial buyers are the earliest audience OpenAI is courting.
Where it lands first
OpenAI’s launch materials point to three settings. Industrial machine vision: a camera over an assembly line flags a defect the moment it appears, instead of after the batch runs. Security monitoring: surveillance feeds receive continuous real-time analysis rather than after-the-fact review, which changes the economics of human-staffed control rooms. Accessibility: a navigation assistant for visually impaired users describes a crossing as it happens, using live video from a phone camera and returning spoken guidance within seconds.
Each of these was technically possible before, but only with custom vision systems built for a single task. The appeal of GPT-5 is that one model covers all of them, and developers describe the capability in near-real-time terms — the seconds-level latency OpenAI cites puts the perception-to-reasoning loop within reach of production use, not just demonstration.
What OpenAI hasn’t said
What OpenAI has not disclosed is nearly as notable as what it announced. The company has not published API pricing or regional availability for the video feature, and it has not said whether the capability will cover Chinese-language scenarios. For developers planning to build on it, that leaves open questions about cost, production latency and which markets will get the service first.
The timing is deliberate. Google’s Gemini line and Anthropic’s Claude models have both added vision, but real-time video streaming is a different register, closer to robotics and live operations than to chat. OpenAI is claiming a first-mover window in a category that is only beginning to form. Streaming video is expensive to process at scale, and the seconds-level latency shown in demos will face real pressure once thousands of developers point live feeds at the API. The company’s caution on pricing and regions suggests it knows the hard part is not the demo.
The capability also opens a route to agentic applications that earlier models could not support. A model that watches a live scene can act on what it sees — calling a tool, sending an alert, updating a record — within the same seconds window it uses to interpret the footage. Developers who build on the video feature describe it as the missing link between perception and action in automation: the system does not just recognize a jam on a conveyor; it can file the report and page the technician in the same breath.
For OpenAI, the launch is also a statement about architecture. GPT-5 is a family, not a single model — the company routes requests through a router that decides how much reasoning a question deserves, and a lighter model handles overflow traffic when limits are reached. The video capability sits on top of that system, which means the same infrastructure that serves ChatGPT’s hundreds of millions of users can carry the new streaming workloads. That is the scale argument OpenAI will need if the video feature is to become a product rather than a demonstration.
OpenAI’s GPT-5 launch is less about a smarter chatbot than about a new input channel: live video. The applications — factory inspection, surveillance, assistive navigation — are concrete, and the latency has reached the point where they are plausible. The open questions are commercial, not technical. Until OpenAI publishes pricing and availability, live video is a capability in search of a market, and the company’s rivals will be watching the stream as closely as its customers.
