Artificial Intelligence Daily edition
ARTIFICIAL INTELLIGENCE GENERATIVE AI
Images 2.5 Moves Visual Generation from Single Renders to an Editing Workflow
OpenAI has updated its image system with lower latency, localised editing and greater continuity between iterations. The practical leap forward is not in producing another eye-catching image, but in preserving identity, composition and decisions as a team works on them.
On 8 September, OpenAI launched ChatGPT Images 2.5 for ChatGPT, ChatGPT Work and Codex across all tiers, along with two models for developers: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. The announcement comes in a market that no longer needs to prove a machine can produce a plausible photograph. The professional question is different: whether the system can take a reference, modify only what is requested, maintain the rest over several rounds, and do so at a time and cost compatible with production. This transition, from surprise to control, is the real substance of the launch.
The stated improvement combines more natural detail, light and textures with superior fidelity to the people and objects in a reference image. OpenAI claims to have reduced latency by up to 50% compared to Images 2.0. More importantly for a catalogue or campaign, the editor promises to alter one element without recomposing the others and to better preserve previous changes in long conversations. This is not a nuance: when a bottle changes shape, a face loses its features, or the background drifts when correcting a colour, each iteration invalidates the previous one and multiplies the need for human review.
Sketch, Templates and Comments: The Prompt Is No Longer Private
The interface attempts to turn that control into a shared process. Sketch allows users to draw within ChatGPT and use the stroke as a spatial guide; templates offer starting points for posters or product photography; and comments placed on the image itself pinpoint a correction. A prompt can also be shared with an image so someone else can test the idea with their own references. The prompt thus ceases to be a secret paragraph and instead coexists with sketches, selections and observations. For a team, this makes it easier to explain why a variant exists, although it is not yet equivalent to a change history as precise as that of a design tool.
The promise also has its limits. A more coherent image can still invent text, product geometry, reflections or regulated details. Preserving a face does not grant permission to use it, and reproducing a recognisable aesthetic does not resolve copyright or trademark issues. OpenAI maintains controls on inputs and outputs, C2PA metadata and invisible watermarks, which are useful for provenance but not infallible when the file is captured, recompressed or passed through other tools. The editor must retain originals, consent, the model used and the purpose of each change, as well as checking the final-size result.
Flare, Sunburst and the Cost Per Approved Asset
In the API, Flare is the recommended option for most workflows and prioritises speed; Sunburst reserves more time for jobs that demand fine control. Both are listed at $8 per million input image tokens, $2 for cached input, and $30 for output; text is charged separately. These are per-token rates, not a fixed price per file, so resolution, references and successive edits must be measured against a real-world batch. Compared to GPT-Image-2, the published unit price has doubled, although OpenAI maintains that Flare delivers higher quality with half the latency. In production, what matters is the cost per approved asset, not that of an isolated generation.
A sensible pilot project does not start by replacing an entire production run. It might take fifty already-approved assets, set a master reference, and request three types of change: background, colour and framing. Then, it measures time to approval, percentage of intact elements, text or logo errors, cost per variant, and number of rounds. The rules must be explicit: do not touch the product, do not introduce commercial claims, do not alter a person's features, and stop when a licence is missing. AI accelerates the draft; quality control decides if the piece can be published.
It is now worth watching whether this consistency survives ten or twenty edits, if Sketch and comments integrate with enterprise approval workflows, and how the cost behaves when an application retains many references. Regional availability, the handling of sensitive data, and the persistence of provenance when exporting to other suites will also be relevant. Images 2.5 reduces friction and opens up a more serious workflow, but it does not eliminate creative work. It shifts it towards direction, selection and providing evidence of what changed and what needed to remain the same.
Tags
- OpenAI
- ChatGPT
- Images 2.5
- Generative AI
- Visual production
- C2PA
BOLDERROR Daily edition Rubén Campoy