GPT-5 Multimodal Strategy: OpenAI Bets on Images, Exits Video

The launch that didn't settle much

The headline "OpenAI unveils GPT-5" is now more than a year old, and it reads differently today than it did in August 2025. GPT-5 launched on August 7, 2025, during a livestream. It reached ChatGPT and Microsoft Copilot users and developers using the OpenAI API at the same time.11 When it shipped, it scored at or near the top of benchmarks covering math, programming, finance and multimodal understanding.11 It has since been succeeded by GPT-5.1.11

The question for anyone following AI video generation, text-to-video models and AI image generation is no longer whether GPT-5 was a big step. The question is what OpenAI built on top of it. The answer is a split. Over the past year OpenAI has put heavy investment into image generation and pulled out of consumer video. That says more about the economics of multimodal AI than any benchmark from launch day.

What "multimodal" meant at launch

The main idea behind GPT-5 was consolidation. Romain Huet, OpenAI's head of developer experience, described it as merging the reasoning gains of the o-series with the multimodal gains of the GPT series into one system.19 Before that, users had to choose between models, for example GPT-4o for multimodal work and o3 for hard reasoning. GPT-5 was meant to remove that choice.19 It did this with dynamic routing: simple prompts went to fast pathways, and harder problems triggered deeper "thinking."16

On understanding images and other inputs, the reported results were strong. Wikipedia's entry, citing AIMultiple, describes GPT-5 as natively multimodal, meaning it was trained on several modalities from the start rather than attached to separate vision models.11 One technical analysis reported an 84.2% score on the MMMU multimodal benchmark.13 Product coverage this year describes a consumer app that handles images, charts and PDFs in one conversation, with voice mode also running on GPT-5.14

The launch design also ran into trouble. Early users reported inconsistent performance and confusing interface changes. Many were also unhappy that GPT-4o had been removed. OpenAI brought the older models back and replaced the automatic routing with settings users choose themselves.17 The unified, decide-for-you model did not survive contact with users in its original form.

Where the coverage diverges: could GPT-5 generate?

The launch coverage disagrees most sharply on output, not input. One widely shared report said GPT-5 included video generation and called it fully multimodal.20 A more careful technical review said the reverse: GPT-5 can understand images but cannot generate images or audio, and that work is handled by separate models.13 Another summary listed text, images and code as things GPT-5 could produce.16

The technical review's version is the one that fits what happened next. If GPT-5 had shipped with built-in video generation, OpenAI would not have needed a standalone Sora app weeks later, and it would not have needed a separate line of image models. Both things happened. In practice, "multimodal" at OpenAI has meant a reasoning model that reads many kinds of input, paired with specialized generators for pictures and video. The claims about video generation at launch look like hype that got ahead of the product.

The image side: a clear winner

On images, OpenAI has done well. GPT Image first appeared in March 2025 as the successor to DALL-E. It went viral for Studio Ghibli-style pictures.1 GPT Image 1.5 followed in December.1 Then on April 21, 2026, OpenAI released GPT Image 2, its first image model that reasons before it generates.58 The consumer version, branded ChatGPT Images 2.0, reached every ChatGPT plan the next day.7

The reported gains are practical rather than flashy. Coverage keeps pointing to text rendering, the long-standing weakness of image models, as the biggest improvement, including reliable output in non-Latin scripts.78 One developer review said the jump over GPT Image 1.5 is smaller for photorealistic images with no text, while menus, mockups and infographics benefit most.8 That matches where business demand for AI images is.

The leaderboards back this up, with a caveat. One September roundup puts GPT Image 2 first on the Artificial Analysis text-to-image arena at an Elo of 1,339. It describes that as the widest first-to-second gap the arena has recorded.10 The same piece notes that another blind-test ranking gives a much lower score, around 1,177, while still placing OpenAI first.10 It also says GPT Image 2 fell to second on a separate image-editing arena after Reve 2.1 took the lead there.10 The fair reading is that OpenAI leads clearly on generation and is competitive, but not dominant, on editing.

OpenAI has kept moving. GPT Image 2.5 arrived on September 8, 2026. It adds a Sketch feature that turns drawings into images and is reported to cut generation latency in half.1 It comes in two API versions: Flare, built for speed, and Sunburst, built for detail.1 The official API reference lists both, with support for arbitrary resolutions and transparent backgrounds.2

The product line has also been cut down. DALL-E 2 and DALL-E 3 were retired on May 12, 2026.5 OpenAI then said gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest will shut down on December 1, 2026, with all three moving to gpt-image-2.9 Some developers who moved to the mini model after the DALL-E retirement now face a second forced migration within about six months.9 That cost is real, and it shows how quickly OpenAI is willing to retire products it has stopped prioritizing.

The video side: a retreat

Video went the other way. Sora moved out of research preview in December 2024 as a standalone product, and OpenAI admitted at the time that it often produced unrealistic physics.21 A second-generation model with audio and better physics came out in September 2025, along with a social-style Sora app.26 The app topped its App Store category within a day.26 In December, Disney signed a three-year licensing deal to bring its characters to Sora and pledged a $1 billion investment.26

Then OpenAI reversed course. Reuters reported on March 10, 2026, citing The Information, that OpenAI planned to bring Sora into ChatGPT.24 Two weeks later, on March 24, OpenAI announced it was closing Sora.23 The app and web product went offline on April 26, and the Sora 2 API was scheduled for removal on September 24.25 According to people familiar with the matter, the Disney deal is no longer going ahead.26 Disney said it respected OpenAI's decision to leave the video generation business.30

The reporting agrees broadly on why, though it puts the emphasis in different places. Compute is the common thread. OpenAI said it had to make trade-offs on products with high compute costs.30 Late in 2025, Sora's head had already limited how many videos users could generate because chips were in short supply.26 One academic commentary cites The Wall Street Journal as reporting that Sora lost about $1 million a day.22 Coverage also points to competition from Anthropic, which skipped image and video entirely to focus on text and code.26 Forbes argued that safety and deepfake worries were real but probably not the main reason, describing a compute-hungry social video app as a luxury next to OpenAI's coding, enterprise and robotics priorities.28 The New York Times connected the move to preparations for a possible IPO.23

What this means for text-to-video

OpenAI has not abandoned video research. The company says the Sora team is now working on world simulation to help train robots.27 That suggests OpenAI sees video models mainly as tools for teaching machines how the physical world works, not as a consumer product.

The Sora shutdown is not evidence that text-to-video is failing as a category. Runway, Kling, Google Veo, Pika and Luma are still active. Veo 3's native audio and its distribution through YouTube give Google a significant advantage.25 CNN noted that Google had drawn attention with recent improvements to its video model just as OpenAI was pulling back.30

This reading of the past year: GPT-5's real legacy is a reasoning model that can read almost any input. It is the foundation for an image business that now leads the benchmarks. Video showed the limits of that strategy. Generating video costs far more than generating text or images, and the revenue did not cover it.22 For now, the multimodal race splits two ways. OpenAI is betting on images that sell, and rivals like Google are taking over the video market OpenAI left.