OpenAI Images (gpt-image)

OpenAI Images (gpt-image) is OpenAI's hosted text-to-image and image-editing API, first released in April 2025 as gpt-image-1, a natively multimodal model built on the same GPT-4o-class image generation that went viral in ChatGPT. Unlike diffusion-based competitors, it is an autoregressive multimodal model that accepts both text and image inputs and produces image outputs, which gives it standout strengths in instruction-following, accurate in-image text rendering, and world-knowledge-aware composition. The product line has since iterated quickly: gpt-image-1-mini (a cheaper, faster variant), gpt-image-1.5 (December 2025, roughly 20% cheaper and a notable quality/text-rendering jump), and gpt-image-2 (the current flagship). All are reachable through the dedicated Images API and the conversational Responses API, the latter enabling multi-turn, high-fidelity iterative edits.

The target user is a developer or product team embedding generation/editing into an app rather than an end-user buying a creative seat. That positioning is validated by marquee launch partners already wiring it in, Adobe (Firefly/Express), Figma, Canva, Wix, Airtable, Instacart, and GoDaddy. Where it wins: best-in-class prompt adherence and legible text (it consistently ranks at or near the top of Artificial Analysis's Text-to-Image Arena), strong multi-image reference editing, transparent-background and format/compression controls, tunable moderation (auto/low), and C2PA provenance metadata baked into every output. Pricing is genuinely usage-based per token, working out to roughly $0.02 / $0.07 / $0.19 per low/medium/high-quality square image on the original gpt-image-1, with later models cheaper.

Where it loses: latency is the dominant complaint, high-quality generations routinely take 30–60 seconds and the API can time out around 180 seconds, which is painful for interactive UX, though streaming partial images and the mini/batch tiers soften this. It is pricier and slower than open-weight diffusion stacks you can self-host, the original model had a noticeable yellow tint, transparency isn't supported on the newest gpt-image-2, and content moderation can produce false positives. Reliability is generally strong (status-page components mostly 99.8–99.99%) but image generation has had recurring incident clusters in 2025–2026. Net: the default choice when you need reliable instruction-following and text-in-image and are willing to pay OpenAI per token and tolerate multi-second latency.