ChatGPT Images 2.0 Adds Web Search as OpenAI Exits AI Video

What OpenAI launched

On April 21, 2026, OpenAI released ChatGPT Images 2.0. The new image generator runs on a model called GPT Image 2, and its headline feature is that it can look things up on the web before it draws.11 When a user picks one of ChatGPT's thinking models, the generator can pull in information from the web, make visual explainers from uploaded files, and work out the layout of an image before it renders anything.11 OpenAI also says thinking mode can check its own output and produce several distinct images from a single prompt.20

The rest of the update is about practical control. With thinking turned on, the model can make up to eight images at once and keep the same characters, objects and style across all of them. OpenAI suggests uses like manga pages, sets of social graphics, or design plans for every room in a house.11 Output now goes up to 2K resolution, with aspect ratios from a wide 3:1 to a tall 1:3.11 OpenAI also claims big improvements in rendering text in Japanese, Korean, Chinese, Hindi and Bengali.11 Getting non-Latin text right inside an image has long been a weak point for OpenAI's image tools.26

This was OpenAI's third image model. It followed the original ChatGPT Images in March 2025 and Images 1.5 that December.14 Sam Altman did not undersell it. During the launch livestream he compared the jump to going from GPT-3 to GPT-5 in one step.20

Where the reporting agrees and where it splits

The outlets agree on the main facts: web access, reasoning before rendering, batches of up to eight images, 2K output and better multilingual text.111417 They differ on who gets what.

The Verge reported that thinking features were limited to Plus, Pro, Business and Enterprise subscribers, while the core improvements reached everyone.11 PCMag listed only Plus, Pro and Business as having the advanced thinking outputs.14 One aggregator summary went further and described the whole launch as subscriber-only.16 OpenAI's own release notes settle it. Images 2.0 is on every ChatGPT plan, and "images with thinking" is on every paid plan when a Thinking or Pro model is selected.24

A bigger gap, and an under-reported one, is what "pulling from the web" actually means. Gizmodo quoted OpenAI's claim that the model can search for real-time information.20 PCMag pointed out that the model's own knowledge stops in December 2025.14 These claims don't contradict each other, but together they set the limits of the feature. Web search is a retrieval layer added on top of a model whose built-in knowledge is fixed. That makes it useful for grounding an image in current facts. It does not guarantee the model understands those facts. Pre-release testing showed why that matters. Images circulating from the testing period included a world map with invented countries and Nairobi placed in Saudi Arabia.20 Grounding is meant to reduce that kind of error. How well it does so is still an open question.

Why web grounding matters for multimodal AI

The most important change in Images 2.0 is architectural, not visual. OpenAI's image generation now works less like a single model call and more like an agent: it plans, searches, generates and then checks itself.12 Some coverage called this a new benchmark for the industry because it moves past static training data.13 That claim is plausible, but OpenAI did not get there out of nowhere. A year earlier, its o3 and o4-mini reasoning models could already chain web search, file analysis and image generation inside ChatGPT.24 Images 2.0 builds that same tool use into the image pipeline itself.

The benchmarks suggest the approach worked. A September 2026 roundup reported GPT Image 2 at an Elo of 1,339 on the Artificial Analysis image arena. By that account, it held the largest lead between first and second place the leaderboard had ever recorded, with Microsoft's MAI-Image-2.6-Preview in second.23 The same analysis credited the model's "think before it draws" design for that lead.23 At launch, PCMag noted that Google's Nano Banana Pro had won its most recent flagship comparison.14 If the arena numbers hold up, Images 2.0 reversed that result.

OpenAI has kept iterating. GPT Image 2.5 arrived on September 8, 2026, with sketch-to-image and about 50% lower latency. It was offered in the API as two versions: Flare, built for speed, and Sunburst, built for detail.21 OpenAI's developer documentation now lists both as its specialized image models.22 Coverage does not agree on the naming. One tracker treats ChatGPT Images 2.5 as a consumer brand that still maps to GPT Image 2 in benchmarks.23 Wikipedia describes 2.5 as a separate model release.21 For anyone deciding what to use, the practical point is the same: OpenAI's image stack is changing every few months.

The competitive pressure behind it

The launch was a direct response to Google. Gizmodo framed the release as OpenAI trying to win users back from Nano Banana Pro. It also reported that Google's late-2025 image and Gemini 3 launches had prompted OpenAI to declare a "code red."20 Other coverage, citing Fortune's reporting, said Google's Nano Banana releases and user growth pushed OpenAI into a company-wide effort.12 The Verge listed Microsoft's MAI-Image-2 alongside Google's tool as part of a crowded field.11

The business stakes are clear. Gizmodo argued that a viral image moment, like the Studio Ghibli wave set off by GPT-4o's image generation, could help push ChatGPT from the 900 million weekly users it reported in February toward a billion. That would make a good story for a possible IPO.20 The first image model shows how much a hit can matter. OpenAI said more than 130 million people made over 700 million images in its first week.21

The video retreat makes images more important

The image push looks different alongside what OpenAI did with video. On March 24, 2026, about four weeks before Images 2.0, OpenAI announced it would shut down Sora.6 The app closed on April 26, 2026, and the API was turned off on September 24, 2026.6 OpenAI's deprecations page reportedly listed no recommended replacement.10 OpenAI gave no official reason. Reports pointed to compute shortages and costs: Sora's users peaked at around a million and then fell below 500,000, while the service was estimated to cost about $1 million a day to run.6 Gizmodo tied the shutdown directly to OpenAI tidying its finances ahead of an IPO.20

That changes how to read Images 2.0. It is not just an upgrade. It is now OpenAI's main visual-generation product. Multi-image consistency, with recurring characters, manga sequences and storyboards, covers some of the storytelling ground that video used to cover, and still images are far cheaper to make. My reading is that OpenAI decided coherent sequences of still frames, backed by reasoning, give it better returns than paying for video compute.

Competitors have taken the space OpenAI left. One October 2026 analysis described the top of the text-to-video-with-audio leaderboard as split between Google and Chinese models from Alibaba, ByteDance, MiniMax and Kuaishou.1 Google told developers to move from Veo 3.1 Preview to Gemini Omni Flash, which makes 3- to 10-second clips at up to 4K and lets users edit video through multi-turn conversation.10 Google describes Omni as "Nano Banana for video," and it is replacing Veo in the Gemini app.5 Kuaishou released Kling 4.0 on October 15, 2026, with native 30-second generation.10 Synchronized audio has also become standard: five of the seven leading alternatives reviewed in one comparison generate sound natively.4 Not every guide has caught up. At least one still describes Sora as a 60-second leader available in ChatGPT Plus.2 That shows how quickly this market is outrunning the people who cover it.

OpenAI has put its audio work into ChatGPT itself, not into a standalone media app. In July 2026 it launched GPT-Live-1 for ChatGPT Voice. The model can use web search and memory and work with text and images in the same conversation.24 Taken together with Images 2.0 and the GPT-6 "Intelligent UI" update, which adds interactive charts and diagrams to answers, the strategy is consistent.27 OpenAI is building generation, search and reasoning into one assistant instead of running separate creative products.

The realism problem

The obvious downside is that the images are more convincing. Business Insider noted that OpenAI's own demo images included fake screenshots and photos of physical magazines. Those are exactly the kinds of images that make it harder to tell real from fake.19 OpenAI researcher Gabriel Goh said photorealism was the style he was most excited about.20 NPR's Here & Now raised the same concern: as AI-generated images get more realistic, telling what is real gets harder.15

On copyright, OpenAI told Business Insider that it blocks copying the styles of individual living artists but allows broader studio styles.19 The EFF's Mitch Stoltz said a resemblance at the level of a style or idea is generally not copyright infringement. He added that the wider social problem is how easy and widely available such replicas have become.19 Web retrieval makes this sharper. A model that can fetch reference images in real time could produce near-copies more easily, though none of the coverage tested this directly.

The bottom line

Images 2.0 is a real step forward. Reasoning and web search have measurably improved accuracy, consistency and text rendering, and the benchmark lead backs that up.23 Its larger significance is strategic. Having dropped Sora, OpenAI is betting that grounded, multi-frame still images inside one multimodal assistant are worth more than costly video.610 Rivals now control the video leaderboards.1 Whether that bet pays off depends on whether coherent image sequences can hold onto users who increasingly expect motion and sound.