ChatGPT Images Grows From GPT Image 1.5 to 2.5 as Sora Shuts Down

A December launch that set the pace

When OpenAI announced "the new ChatGPT Images" on December 16, 2025, it looked like a routine model update. Looking back from late 2026, it marked the start of a broader shift in how the company handles media generation. Still images became a core part of ChatGPT that gets upgraded on a regular schedule. Standalone video was dropped. Voice moved into the developer platform as infrastructure for AI agents.

The December release added a new flagship image model to ChatGPT for all users and made it available in the API as GPT Image 1.5. OpenAI said it could make precise edits while keeping the rest of an image intact, and could produce images up to four times faster than before.1 The company also added a dedicated Images area to the ChatGPT sidebar with dozens of preset filters and prompts, which it promised to refresh as trends change.1 API pricing for image inputs and outputs fell 20% compared with GPT Image 1, and OpenAI aimed the model at marketing teams and online retailers who need product catalogs built from a single source image.1

The timing was competitive. One review noted that OpenAI shipped the update less than a month after Google released Nano Banana Pro, and called it OpenAI's answer to that model.3

How the speed claim held up

OpenAI's announcement and outside testing told slightly different stories. The company said images were generated up to four times faster.7 A hands-on review found that full new images were clearly faster but did not reach that figure. The reviewer concluded that most of the gain came from regenerating only the edited parts of an image, and that the same partial regeneration explained most of the lower API cost.3

Early users were split. Some said the yellow tint in earlier ChatGPT images had faded and that the two leading models were now roughly level.6 Others disputed OpenAI's claim about preserving images, saying uploaded photos came back less consistent and with plastic-looking skin.6 One photographer liked the editing but said the output resolution was too low for professional work.6 Another commenter pointed to price, putting Nano Banana Pro at about 13 cents per image and OpenAI's model at about 3 cents.6

OpenAI admitted some of these limits itself. It said results were still imperfect and kept the earlier image model available as a custom GPT.1 Reviewers noted that the update got worse at certain art styles even as it fixed problems like premature cropping and the warm color cast.23

On balance, December 2025 was a clear win for editing and control rather than raw image quality. OpenAI's own wording, which stressed instruction following and preserving details, supports that view.7

Three releases in nine months

What followed shows how quickly OpenAI now updates its image model. In April 2026 it released ChatGPT Images 2.0, with the API model gpt-image-2. Coverage described it as OpenAI's first image model that reasons before it draws.10 When a reasoning model is selected, it can search the web, produce several distinct images from one prompt, and check its own work. One report put the limit at eight coherent images per prompt, at up to 2K resolution through the API.10

Independent testers praised its text rendering. TechCrunch said a generated restaurant menu could pass for the real thing.4 WIRED agreed on English text but found that a Chinese-language poster was full of half-gibberish characters, so multilingual claims still need checking.5 OpenAI would not tell reporters what kind of model powered Images 2.0.4 That leaves the architecture uncertain, even though reference summaries describe the GPT Image family as autoregressive rather than diffusion-based.2

Reports also differ on when developers could use it. One account said the gpt-image-2 API went live on launch day.14 Another said developers would not get access until early May.11 Either way, the release came with a clean break from the past: DALL-E 2 and DALL-E 3 were scheduled for retirement on May 12, 2026.12

Then on September 8, 2026, OpenAI released Images 2.5. It added a Sketch feature that turns drawings into finished images, cut generation time by up to half, and split the API into two models. Flare is the faster option and Sunburst the more detailed one.216 Clean-up of older models continued. One account reports that gpt-image-1 was shut down on October 23, 2026, and that the chatgpt-image-latest alias ends on December 1, 2026.13

The gaps between releases are telling. GPT Image 1 arrived in March 2025, a cheaper Mini version in October, 1.5 in December, 2.0 in April and 2.5 in September.2 One tracker noted that OpenAI has shipped a new image model about every four months.17 Images are now a product that OpenAI improves continuously.

Video went the other way

OpenAI's video story ran in the opposite direction. According to several reports, the company said in March 2026 that it would discontinue Sora. The web and app versions shut down on April 26, and the API followed on September 24.2526 One outlet said OpenAI blamed compute costs and a move toward enterprise products, and that no successor is planned.27 Another said OpenAI's deprecation notice for the Sora API names no replacement.22

That matters because Sora was the product that brought AI video to a mass audience.25 Early in 2026, some comparisons still gave Sora 2 an edge on cinematic quality while crediting Google's Veo 3.1 with better audio sync and true 4K output.24 By October, rankings looked different. One survey of the text-to-video-with-audio leaderboard found Google's Gemini Omni Flash at the top, with Chinese models from Alibaba, ByteDance, MiniMax and Kuaishou close behind.20 Another analysis found that the best Chinese models sit at the low end of the price range while Veo 3.1 costs about ten times more.25

Buying guides have adjusted. They now recommend Veo, Kling, Seedance or Runway for video and treat GPT Image 2 or Nano Banana Pro as tools for still frames, such as the first frame of a clip.21 One multi-model platform offers a GPT image-to-video pipeline that pairs OpenAI's still images with motion from other companies' video models.22

The likely explanation is that OpenAI chose to put its compute into images, where usage inside ChatGPT is huge. The first GPT image model drew more than 130 million users who made over 700 million images in a single week.2 Video is expensive to generate and has a crowded field of rivals, and OpenAI appears to have decided it was not worth the cost.

Voice becomes infrastructure

Voice took a third path. On May 7, 2026, OpenAI released three audio models through its API: GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper. Reuters reported early test customers including Zillow, Priceline and Deutsche Telekom.29 Some coverage dates the launch to May 8, which probably reflects time zones or when the announcement went out.3738

OpenAI called GPT-Realtime-2 its first voice model with GPT-5-class reasoning. It has a 128K context window, up from 32K, and the Realtime API came out of beta at the same time.31 The model handles speech in and speech out in one step, with reasoning happening inside the audio loop instead of passing through separate transcription and speech stages.35 Developers can set reasoning effort at five levels, and the model can say short phrases such as "let me check that" so callers don't sit in silence.30 On OpenAI's own Big Bench Audio test, the high setting scored 96.6%, up from 81.4% for the previous version.31

Not everyone was impressed. One business-focused analysis called the release incremental rather than revolutionary and warned that output costs of $64 per million audio tokens add up quickly.33 One reviewer pointed out that the voice mode in ChatGPT had until then lagged well behind OpenAI's text models.30 OpenAI followed up in July with GPT-Realtime-2.1 and a cheaper mini version, and said latency had dropped across its voice models.32

What it adds up to

Taken together, OpenAI's media strategy now has three clear tiers. Images are a consumer feature inside ChatGPT, updated every few months and expanding from editing to reasoning to sketching. Voice is mainly a developer product built for customer service and other agent tasks. Video, which once made the biggest headlines, has been left to Google and Chinese companies.

The December 2025 launch set this pattern. Its focus on precise edits, a dedicated workspace and lower API prices, rather than resolution or cinematic polish, showed what OpenAI would prioritize.13 The risk is that "multimodal" now means images, text and voice at OpenAI, while rivals offer creators a single toolchain that runs from still image to finished video.23 For now, OpenAI appears willing to accept that trade-off.