ChatGPT Images 2.5: faster, but the safety numbers are not significant
Days after the GPT-6 Astra announcement, on 8 September 2026 OpenAI introduced ChatGPT Images 2.5, the new image generation model available in ChatGPT, ChatGPT Work and Codex. For developers the offering splits in two on the API: GPT-Image-2.5 Flare, set as the default and built for speed (with the documented snapshot gpt-image-2.5-flare-2026-09-08), and GPT-Image-2.5 Sunburst, aimed at precise control and editing. The update brings features such as the Sketch tool, which lets you draw directly inside ChatGPT to steer generation, templates for posters and merchandising, and pinpoint comments on parts of an image for targeted edits. Against a claimed volume of more than 3 billion images created every week across the platform and the API, the vendor reports latency down by up to 50% compared with Images 2.0 in the best case: a figure quoted in the official announcement that comes straight from the company, not from independent third-party benchmarks.
On API costs, the price list sets 5 dollars per million tokens for text input (1.25 when cached), 8 dollars for image input (2 cached) and 30 dollars per million tokens for output, identical for Flare and Sunburst. Because token counting differs from the previous model and there is no per-image calculator, the actual amount for a single generation cannot be worked out in advance. On safety, the automated adversarial evaluation reported in the system card shows a share of unsafe content shown to the user of 1.09% for Sunburst and 1.41% for Flare, against 1.64% for Images 2.0. Yet OpenAI's own document explicitly admits that none of these differences meets the statistical significance threshold of p below 0.05, and notes that the tests rest on a fixed set and that automatic labels can contain errors.
The Deployment Safety Hub report also acknowledges that the model's heightened realism can make for more convincing deepfakes of real people, places or events. To counter these breaches of its guidelines, OpenAI applies filters at both prompt and image level, alongside C2PA metadata and Google DeepMind's invisible SynthID watermarking. As for the Preparedness Framework, neither Flare nor Sunburst crosses the Bio High or Cyber High risk thresholds, though the company is keeping the mitigations designed for high biological capability in place as a precaution.
Three decimal points of difference on an internal test do not make a model safer, and OpenAI is the first to say so. The main news remains the shorter waiting times the company reports, up to 50% in the best case, leaving open the question of how the filters will hold up against images that are ever harder to tell from reality.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- I opened the official system card at deploymentsafety.openai.com (an OpenAI domain) and took from it the safety percentages, the note on statistical significance and the passages on realism and deepfakes. The API documentation at developers.openai.com confirms the per-million-token prices, the endpoints and the snapshot dated 2026-09-08. As independent confirmation, 9to5Mac (8 September 2026) and Unite.AI: the date, the names of the two API models, the user-facing features and the latency claim. The announcement page on openai.com returned HTTP 403 to our automated access: no fact in this article rests on it. I dropped the 4K output claim, which appears only on paid press-release wires.
- Incertezze
- The drop in unsafe content shown (from 1.64% to 1.09%/1.41%) is described by OpenAI itself as not statistically significant: one can say the new model is no worse on the test used, not that it is safer. All the safety figures come from an internal adversarial test on a fixed set, with automatic labelling the company admits can be wrong, and no independent verification exists. The 50% latency cut is a vendor claim, not a third-party measurement; the speed comparisons quoted in the announcement come from commercial partners. With no per-image calculator and token counting that differs from GPT Image 2, the real cost of a single generation cannot be worked out in advance. Claims of 4K output circulating on paid press-release sites have no official confirmation. How well C2PA and SynthID survive recompression, screenshots or metadata stripping cannot be verified.
- Perché pubblicarla
- It is a flagship release with complete, verifiable technical and safety documentation, so it can be reported without relying on rumour. Above all it is a textbook case in reading numbers critically: the company publishes an improvement in its safety metrics and, in the same document, states that the improvement is not statistically significant, while admitting that greater realism makes deepfakes more convincing. Readers need someone to separate what is measured from what is merely claimed, especially with the AI Act's transparency duties on synthetic content coming into force. Image generation, moreover, is not yet covered in our archive.