Bulk-generate era for the
image API — GPT Image 2.0 gets a 50% Batch cut.
The image API is moving from "one call, one image" to "queue a stack, receive a stack for half the price." OpenAI just opened Batch API on GPT Image 2.0 (the successor to DALL-E), pricing async bulk generation at 50% off the standard rate — right after the legacy DALL-E 2 / 3 API sunset on May 12.
The dial moved from
"cheap over fast" to "fast over cheap"
Two things have shifted the image-API landscape in the past two months. First, on May 12 OpenAI sunset the DALL-E 2 and DALL-E 3 API endpoints via deprecation. Second — and this is today's move — its successor, GPT Image 2.0 (codename gpt-image-2), just gained Batch API support.
Image APIs have been a synchronous game: one call, one image, cents per shot, stacked linearly. With this update, any workload that can tolerate up to 24-hour turnaround can now run the same number of images at 50% of the standard rate. It is not a discount so much as a second lane on the same road: one prioritises latency, the other prioritises cost.
| Sync requests (existing) | Batch API (today) |
|---|---|
| One call → one image, returned instantly | JSONL file in, results back within 24h |
| Latency: seconds, small | Latency: up to 24 hours, large |
| Standard rate (100%) | 50% off standard |
| Best for chat UIs, real-time UX | Best for bulk catalog / dataset / brief generation |
Stop counting by image;
start counting by the thousand
Three numbers sum up this story. Against the standard synchronous endpoint, Batch runs at 50% off. The queue promises a 24-hour ceiling on turnaround, returning finished jobs progressively as a JSONL file. And in the background, May 12, 2026 was the day the legacy DALL-E 2 and 3 endpoints stopped accepting calls at all. Any code that has not migrated is not calling a live image API today.
On a single image the difference is a few cents. On a thousand it starts to matter. Teams running product catalogs, ad A/B variants, or synthetic training data at daily volumes of hundreds to thousands of images can now shave double-digit percentages off monthly spend — closing in on a fractional engineer's payroll for exactly zero code change beyond routing.
Who this actually helps
The impact isn't uniform — it lands in four distinct places.
Bulk-generation devs and data teams
Synthetic data, e-commerce catalogs, ad variants, video preview stills. Any pipeline that pushes hundreds of images a day can realistically expect double-digit percentage monthly savings by re-slotting jobs into a 24-hour batch — often with no logic change beyond the endpoint.
SaaS / product integrations
Anywhere the user can wait — nightly report cover generation, weekly digest artwork — Batch becomes the default. In-chat real-time image generation still belongs on the sync endpoint. The clean answer is a two-lane router, and this update is the trigger to build it.
Code still calling legacy DALL-E
Separate from Batch, but urgent: per OpenAI's deprecation ledger, the dall-e-3 and dall-e-2 endpoints stopped serving on May 12. Migrate to gpt-image-1 or gpt-image-2 first, then think about batching.
Individuals and ChatGPT image users
Batch is an API-only lever, so this does not touch the images you draw inside ChatGPT itself. That said, cheaper underlying inference tends to reach end users eventually — as quota, quality, or both.
The image API stops being
an interactive tool priced by the shot,
and starts being factory equipment priced by the pallet.
Three limits
worth planning for
First, the latency trap. A 24-hour ceiling is not an SLA. In heavy queues, expect jobs to sit near the ceiling — so a batch scheduled for tomorrow morning's deadline needs a retry path that falls back to sync when 25 hours pass without a response.
Second, moderation still applies. Batch does not bypass content policy. Rejections come back as errors mixed into the JSONL, so plan for perhaps a small percent of any large batch to fail. Provision an error column in whatever table you write results into, and design the downstream step to walk past holes rather than assume a complete grid.
Third, in-flight cancels are limited. Once a batch is submitted, most of it will run. Halving the unit price also halves the pain of a bad prompt in absolute terms — but the mistake now materialises as a thousand wrong images instead of one. Always front a batch with a 10-image dry-run queue; the operational fix is that simple.