Gemini Tops ChatGPT in Vision Tests as OpenAI Exits AI Video

Seeing is now its own contest

AI assistants are no longer judged mainly on how well they write. A newer question is how well they understand what a phone camera points at, and how well they can make video. On the first question, most of this fall's reporting favors Google's Gemini over OpenAI's ChatGPT, though the evidence is mixed. On the second, OpenAI has stopped competing.

CNET's latest head-to-head ran both chatbots through a series of visual tests. It said there was a clear winner on multimodal tasks, especially image recognition, and that winner was Gemini.1 In the clearest example, Gemini transcribed messy handwritten notes word for word, identified where the text came from, and described features of the handwriting. The reviewer called the gap with ChatGPT "night-and-day."1 ChatGPT did win one round, a Venus flytrap identification. But Gemini was more consistent and correctly identified sources in both the notes test and a movie-art test, which was enough for CNET to pick Gemini overall, "for the most part."1

The verdict matters because it comes as both companies make very different bets on multimodal AI. For now, Google is the only one of the two shipping on every front: seeing, listening and generating video.

Where reviewers agree, and where they don't

Other reviewers mostly agree with CNET, but not completely. One hands-on review found that both tools identified objects correctly, while Gemini gave more detail, especially when reading text inside images. It gave Gemini the win for OCR and image context.2 Several broader comparisons call Gemini the stronger multimodal system because it handles images, video, audio and PDFs natively in a single prompt.47 One analysis says Google's pitch is now "native multimodality" while OpenAI's is "agentic intelligence," meaning software that does work on your computer for you.10

The split shows up on handwriting, the very test where CNET saw the biggest gap. G2's tester found the reverse. ChatGPT had a sharper eye for detail in image analysis, while Gemini skipped one section of a handwritten page, repeated another, and spent more time describing the paper than transcribing the text.6 A separate guide also says ChatGPT usually does better with messy paperwork and handwriting, and that Gemini does better with clean, printed documents.3 That guide adds that ChatGPT tends to read complex or unusual charts and technical screenshots more reliably, while Gemini is faster at recognizing real-world objects such as plants, products and landmarks, which it credits to Google's long history with image search and Lens.3

Our reading is that these differences have more to do with the task than with a single quality ranking. Gemini's advantage looks real where identifying something depends on outside knowledge, like matching an image to a source, a place or a product. CNET's handwriting result supports this: the reviewer suggested that Gemini recognizing where the notes came from may have helped its transcription.1 When the job is close reading of a messy page with no outside clues, the results are much less clear.

The specs don't fully agree either. One comparison says ChatGPT has no native video or audio understanding.4 Another says paid ChatGPT handles about 10 minutes of native video input, compared with up to two hours for Gemini's top tier.10 A third says ChatGPT Plus subscribers can upload a video file on web and desktop and ask questions about it.20 What they share is a gap in scale. Gemini can take in far longer media, and Google markets that as a central feature.

The live camera gap OpenAI created for itself

The sharpest difference is in real-time vision, and OpenAI created it. In July, ChatGPT made GPT-Live its default voice mode. It is a full-duplex system, meaning it can listen while it speaks, and it hands harder questions to a larger model in the background.18 It launched without camera input or screen sharing, and OpenAI said only that those features were coming "soon."11 Users who want ChatGPT to see while they talk have to switch back to the older Advanced Voice Mode, which still offers video and screen sharing on mobile.1315

Gemini Live, meanwhile, streams camera and screen on Android and iOS on every plan, including the free tier.1314 One voice-assistant roundup says camera and screen support is a category Google has to itself among the four assistants it tested.14 Another sums up the tradeoff: OpenAI bet on conversation, and Google bet on perception.11

This shouldn't be overstated. ChatGPT can still see through its live camera in the older mode, and one October comparison calls live video chat on mobile a tie.20 Independent benchmarks also favor OpenAI on conversation. A full-duplex voice benchmark found that OpenAI's underlying realtime model handled interruptions best, while Gemini Live had the lowest turn-take rate in the test.15 Still, OpenAI's newest and default voice experience can't see. For a feature built around showing an assistant what you're looking at, a hidden fallback mode is a step down.

OpenAI walked away from video generation

The biggest reason the multimodal race has tilted toward Google is that OpenAI pulled out of making video. OpenAI shut down the Sora web and mobile apps on April 26, 2026, and turned off the Sora API on September 24.41 OpenAI's notice didn't give a detailed reason. Later reporting pointed to computing costs of roughly $1 million a day, users falling from about a million at peak to under 500,000, and a shift of resources toward coding and enterprise products.4942 OpenAI's partnership with Disney, which had licensed Disney characters for Sora, was also ending.49 One developer-focused account says OpenAI described the change as moving the Sora team toward world-simulation research for robotics.48 No replacement video model has been announced.48

The timing is striking. Less than a year earlier, independent reviewers were still debating whether Sora 2 or Google's Veo 3 was better. Mashable called Veo 3 the clear winner overall, while giving Sora 2 credit for putting the reviewer's own likeness into video.23 Other testers preferred Sora 2 for physics-heavy scenes, lip-sync and creative control.30 Some 2026 comparisons that still list Sora as an active rival are out of date, which shows how fast this market is changing.8

Google's answer: one model for every kind of input

Google's response is Gemini Omni, a model family announced at I/O on May 19. Its first model, Omni Flash, generates video from any combination of text, images, audio and video.36 The difference from Google's existing Veo line is that Omni can use an existing video as the basis for a new one, and draws on more world knowledge from Gemini's training data, according to Google DeepMind's CTO.36 DeepMind chief Demis Hassabis said Omni combines Gemini with Google's media models, including Veo, Nano Banana and Genie. He called it a step toward a world model that can understand and simulate the world.37 Omni replaced Veo as the default video model in the Gemini app, though Veo is still used in other Google products.35 Every Omni clip carries an invisible SynthID watermark.38

Google has rolled Omni out quickly. The stable API model reached general availability on August 27, at about $0.10 per second of 720p video.31 It can extend scenes in 10-second steps up to 40 seconds, generate the footage between a set first and last frame, and upscale finished work to 4K.32 On September 23, Google made Omni-powered video generation free in Google Vids for anyone with a Google account.3234 Days later, it added Omni to Google Ads' Asset Studio as a free feature for marketers.33 Adobe has also integrated Omni Flash into Firefly.32

What it means

Gemini's visual lead is probably narrower than its biggest fans claim. Head-to-head tests disagree on handwriting and charts, and ChatGPT remains a strong, sometimes better, reader of difficult images.36 On the overall strategy, though, the picture is clear. Google is building one multimodal system that sees through your camera, takes in hours of media, and turns what it understands into video it can put in front of creators and advertisers for free.141033 OpenAI has chosen to focus on conversation, agents and coding, and has stepped back from video for now.1842

For users choosing today, the advice is simple. If the task depends on showing the assistant something live, or on making video, Gemini is the default. ChatGPT's strengths are elsewhere.