DALL-E 3

Read this first: DALL·E 3 is retired

DALL·E 3 no longer exists as a purchasable product, and this review is scored accordingly. OpenAI's API changelog is explicit: the DALL·E model snapshots (dall-e-2 and dall-e-3) were deprecated and removed from the API on May 12, 2026, with OpenAI recommending gpt-image-2, gpt-image-1 or gpt-image-1-mini instead. Any integration still calling the dall-e-3 model ID now fails outright. On the consumer side, OpenAI's release notes announce that the official DALL·E GPT inside ChatGPT was retired on August 30, 2026, with users pointed to ChatGPT Images. The image generation you get in ChatGPT today — Free, Go (US$8/month), Plus (US$20/month), Pro (US$100 or US$200/month) — is served by the GPT Image model family, not DALL·E 3. I am keeping this review up because the name still carries recognition and because teams with legacy DALL·E 3 integrations need the migration facts, but treat the section below as a historical record with a live migration note.

Where it sat

Context for readers who arrived late: OpenAI's image line ran DALL·E (2021, research preview), DALL·E 2 (April 2022, the first broadly available one), then DALL·E 3 (October 2023). DALL·E 2 was the model that made text-to-image a consumer feature; DALL·E 3 was the one that made it reliable enough to put inside ChatGPT for hundreds of millions of users without embarrassing outputs. For roughly two and a half years it was OpenAI's only first-party image model in production, and its deprecation in favour of the GPT Image family marks the point where image generation stopped being a separate product line and became a native capability of the chat models themselves.

DALL·E 3, released October 2023, was OpenAI's third-generation text-to-image model and the first one I considered genuinely usable for production copy-driven work. Its defining trait was prompt fidelity: where other generators of its era treated a text prompt as mood music, DALL·E 3 attempted to obey it literally, including rendering legible words inside the image. OpenAI integrated it into ChatGPT, where conversational iteration — "make the font bigger, now move the plant" — became the workflow, and exposed it through the API at per-image pricing.

What it did

Text-to-image generation at 1024x1024, 1024x1792 or 1792x1024, in Standard or HD quality, with a style parameter (vivid or natural). API pricing at time of its removal: US$0.040 per Standard 1024x1024, US$0.080 for HD 1024x1024 or Standard wide/tall, US$0.120 for HD wide/tall — per-request billing, no charge for content-policy rejections, no batch discount, and a rate limit of roughly 7 images per minute on the standard tier. The API returned a URL valid for 60 minutes or base64, supported n=1 only (one image per call), and offered no transparent backgrounds and no editing endpoint for the v3 model. In ChatGPT, generation was bundled into the subscription, with free users capped at a small daily allowance and Plus users getting substantially more.

My testing notes

I ran DALL·E 3 heavily through late 2023 and 2024 for concept art, marketing mockups and this site's own illustrative needs, so these observations are drawn from sustained use rather than a one-off test.

Two observations from when it was live. First, prompt adherence and typography were the whole product: on poster-style prompts with 8–10 word headlines, DALL·E 3 rendered the words correctly more often than any 2023 competitor, and it reliably followed negative constraints ("no text", "empty table") that other models ignored. That made it the default for quick concept mockups inside ChatGPT. Second, it was slow and cautious: generation ran on the order of 8–20 seconds per image by size and quality, the n=1 restriction forced one call per variant, and the content filter rejected anything near a brand, a public figure or an edgy theme — sometimes the same prompt passed one day and failed the next, which I found maddening for client work with deadlines. A third observation on cost control: the pricing grid had sharp step changes — square Standard was US$0.040 but a wide or tall image doubled to US$0.080 even at Standard quality, and HD wide/tall hit US$0.120. For anything that could live in a square crop I stayed on Standard 1024x1024, and the quality difference to HD was visible mainly in texture detail on close inspection, not at social-media sizes. Finally, the style flag mattered more than the documentation implied: "vivid" pushed saturation and drama in ways that helped posters but made product mockups look synthetic, so I defaulted to "natural" for anything meant to look photographed.

The limits, in detail

It is worth being precise about what DALL·E 3 never had, because these gaps drove most purchase decisions at the time. No image-to-image or inpainting on the v3 API — those lived only in ChatGPT's conversational edit flow or in DALL·E 2's legacy endpoints. No seed parameter for exact reproduction. No batch calls (n=1), which made bulk generation a loop of single requests at 7 images per minute. No transparency, no vector output, no 2K or 4MP option. And the safety layer sat between you and the model at all times: OpenAI rewrote prompts internally before generation, which occasionally changed results in ways you could not see from the API response.

What replaced it

The GPT Image family. For API users, current per-image equivalents at 1024x1024 run roughly US$0.005–0.036 (gpt-image-1-mini), US$0.009–0.133 (gpt-image-1.5), and wider ranges on gpt-image-2 and the token-billed 2.5 models — OpenAI's own pricing page is the authority and these numbers move. Migration is not a drop-in rename: note also that the "hd" quality value itself is deprecated along with the DALL·E snapshots, per OpenAI's API reference. GPT Image models use different quality parameters, bill by tokens on newer variants, support transparent backgrounds and editing, and generally outperform DALL·E 3 on both quality and cost. If you have dall-e-3 calls in production, the cheapest correct move is to migrate to gpt-image-1-mini or gpt-image-1.5 and re-validate your prompt patterns.

Where it fell short, and the verdict

DALL·E 3's limits were: no editing API, n=1 only, a conservative content policy, a 1792px ceiling, and no self-hosting option ever. And now the hard limit: it is gone. Rated 2.2 on the ToolSpotter rubric, because a retired product cannot score on current functionality depth (the 40%) even though the historical core was strong; value (the 20%) now reads as zero purchasable value with replacement models cheaper per image; usability (the 20%) remains fine for ChatGPT users because image generation simply migrated to newer models; OpenAI's ecosystem and documentation are mature (the 10%); and the reliability signal is a clean, well-communicated deprecation with a working migration path (the 10%). If your question is "should I buy DALL·E 3 today," the answer is you cannot — start with ChatGPT Images or the GPT Image API instead. If your question is "was it good," the honest answer is yes: it was the prompt-follower and text-renderer of its generation, its weaknesses (speed, filter conservatism, no editing) have all been addressed by its successors, and the migration OpenAI forced is one of the rare ones where the replacement is cheaper and better, and OpenAI kept the same images endpoint so the plumbing survives. Teams sitting on legacy dall-e-3 code should budget an afternoon for the endpoint change and a round of prompt re-validation, and nothing more.

Sources

• https://developers.openai.com/api/docs/changelog
• https://developers.openai.com/api/reference/resources/images/methods/generate
• https://help.openai.com/en/articles/6825453-chatgpt-release-notes
• https://help.openai.com/en/articles/6825453-how-to-use-dall-e
• https://aiimagecompare.com/pricing/dall-e
• https://aitoolpick.org/blog/dall-e-pricing-2026
• https://tokenmix.ai/blog/dall-e-api-pricing
• https://modellix.ai/blog/openai-image-generation-api1
• https://thetoolsverse.com/tools/dall-e
• https://aiproductivity.ai/pricing/dall-e-3