C2PA provenance does not distinguish Codex, GPT Image 2.5 Flare, and Sunburst outputs

Summary

This issue is about the signed C2PA provenance contract of the generated artifact, not about adding a model selector to Codex.

Images generated through these three routes all contain the same action metadata:

{
  "action": "c2pa.created",
  "softwareAgent": {
    "name": "gpt-image",
    "version": "2.0"
  }
}

Generation route	Requested/selected model	C2PA softwareAgent
Codex built-in image generation	Not exposed by the tool	gpt-image / 2.0
OpenAI Images API	gpt-image-2.5-flare	gpt-image / 2.0
OpenAI Images API	gpt-image-2.5-sunburst	gpt-image / 2.0

The Flare and Sunburst requests used the documented model IDs. The resulting files therefore demonstrate that softwareAgent.version = 2.0 cannot safely be interpreted as the effective image-model version.

Why this is distinct from #43965

#43965 tracks the Codex built-in tool contract:

• no caller-visible model selector;
• no effective-model field in the tool result;
• uncertainty about how the Codex route is configured.

This issue tracks a different layer: artifact-level provenance remains unable to distinguish the models even when the caller explicitly selects Flare or Sunburst through the documented API.

Fixing #43965 by adding a Codex selector or response-only effective_model field would not fix this issue. A downloaded image examined later or passed to another system would still carry the same gpt-image / 2.0 C2PA value regardless of which documented 2.5 model generated it.

#44369 was correctly closed into #43965 because it only reproduced the Codex-side ambiguity. The API-controlled comparison here is the differentiating evidence.

Relevant published contracts

OpenAI's Images 2.5 announcement distinguishes:

• Images 2.5 in Codex;
• gpt-image-2.5-flare in the API;
• gpt-image-2.5-sunburst in the API.

https://openai.com/index/introducing-chatgpt-images-2-5/

The official image-generation guide shows callers explicitly selecting gpt-image-2.5-sunburst through POST /v1/images/generations and the Responses image-generation tool:

https://developers.openai.com/api/docs/guides/image-generation

The C2PA 2.4 specification defines an action's softwareAgent as the software or hardware used to perform the action. Its generator-info-map.version is described as a human-readable product version, not specifically an underlying model ID:

https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html#_softwareagent

https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html#_generator_info_map

This makes gpt-image / 2.0 ambiguous. It may identify a shared media-service software agent or pipeline rather than GPT Image 2 model weights, but OpenAI's image documentation does not define that semantic distinction.

Impact

• Users may incorrectly treat version: 2.0 as proof that GPT Image 2 was used.
• Conversely, a caller who explicitly selected Flare or Sunburst cannot verify that fact from the signed artifact.
• Offline provenance consumers cannot distinguish GPT Image 2, 2.5 Flare, and 2.5 Sunburst outputs.
• Integrations such as gateways, audit systems, and qgrid cannot safely derive an effective model from the artifact and must retain separate request/response metadata.
• The identical value has already caused multiple Codex reports to conflate a C2PA software-agent version with the backend model version.

Requested clarification / change

1. Document what OpenAI's C2PA softwareAgent.name = gpt-image and version = 2.0 identify: the model family, media-service pipeline, claim-generator product, or something else.
2. Explicitly document that this field is not an effective-model identifier if that is intentional.
3. Preferably include a signed, model-specific provenance value for generated images, such as a documented namespaced C2PA action parameter containing the effective model or stable model family.
4. If exposing the effective model in the artifact is intentionally unsupported, document the supported verification/audit mechanism and the relationship between requested model, effective model, and C2PA software-agent version.
5. Keep Codex and API guidance from presenting softwareAgent.version as proof of the model that ran.

Evidence limits

• The observed C2PA value alone does not prove that Codex or the API ran older model weights.
• This report does not claim that Flare or Sunburst selection was ignored.
• The key result is narrower: distinct, explicitly selected API models produce artifacts with the same model-ambiguous signed software-agent identifier.
• The issue is therefore about provenance semantics and auditability, not generation quality or successful image delivery.

Start by reviewing the C2PA evidence and the linked OpenAI image-generation and C2PA specifications cited in the issue, then compare the Codex and API provenance contracts. Done means the semantics of softwareAgent are documented, model-specific provenance or its intentional absence is addressed, and guidance no longer treats version 2.0 as proof of the effective model.