Does ChatGPT Images 2.5 add a watermark? C2PA explained
OpenAI says ChatGPT Images 2.5 uses C2PA metadata and invisible watermarking to identify its images. What that means for files you get through an API.

OpenAI does not describe a visible watermark for ChatGPT Images 2.5. On its launch page it says it continues to use C2PA metadata and invisible watermarking to help identify images made with its tools.
Facts about OpenAI come from its launch page, read 2026-09-29. Facts about Sume come from the Image API docs. The two do not overlap on this topic, and the gap matters if you generate GPT Image 2.5 pictures through Sume.
What does OpenAI say about safeguards?
The page's safety section says ChatGPT Images 2.5 builds on OpenAI's existing safeguards, with checks on prompts and images to help prevent harmful outputs. It then names the two identification methods and points to a system card for the evaluations. It does not give the technical detail of either mark on that page.
| Mechanism | What the page says |
|---|---|
| C2PA metadata | Used to help identify images made with OpenAI's tools |
| Invisible watermarking | Used to help identify images made with OpenAI's tools |
| Visible watermark | Not mentioned on the page |
| Prompt and image checks | Used to help prevent harmful outputs |
What is C2PA metadata?
The OpenAI page names C2PA metadata and invisible watermarking but gives no technical detail on either. Whether a mark survives a re-save, re-encode or screenshot is not stated there, so test it on your own files rather than assume. A C2PA reader run on a real output file shows whether provenance data is present; that check is yours to make, because neither OpenAI's page nor Sume's docs promise it for files you receive through an API.
Do the marks survive when I generate through Sume?
Sume's docs do not say. They describe the result as a Sume-hosted URL in data[].url rather than inline base64, and note that Sume already mirrors generated media. Nothing in the docs mentions C2PA, watermarking or metadata handling for images, so this post makes no claim either way. If you need the marks, inspect a real output file with a C2PA reader before relying on it.
Do I still have to disclose AI images?
A watermark and a disclosure are separate things. Platforms set their own rules. YouTube's help page, read 2026-09-29, says it requires creators to disclose when they use AI to meaningfully alter or generate photorealistic content. Check the platform's own page for each place you publish; YouTube AI content policy walks through that one.
Where does GPT Image 2.5 fit on Sume?
OpenAI's page separates the ChatGPT product from the API models GPT-Image-2.5 Flare and Sunburst. Sume lists them as openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst. See GPT Image 2.5 release date and model ids for the naming.
Sources
Related posts
More in Models
- Claude Sonnet 5.5 or Opus 5.5 for an agent that calls a video tool?
Sonnet 5.5 and Opus 5.5 differ in price and release date on Anthropic's pages. Which facts matter when the tool is Sume's video generation, and what stays same?
- How to colorize a black and white video with AI
Colorize a black and white video with an AI video-to-video edit: send the clip and a prompt naming the colors. The model invents them.
- ElevenLabs Eleven v4 API: what it is and how to reach it
Eleven v4 and v4 Turbo launched September 28, 2026 on ElevenLabs' own API. What ElevenLabs says they do, and which TTS ids Sume's router lists.
- FLUX.2 dev commercial use: what the model card says
FLUX.2 [dev] weights carry the FLUX non-commercial license, yet the card says outputs can be used commercially as that license describes. The exact wording.
Written by Sume