OpenAI Image Stack — when and how should I migrate?

Decide whether to migrate image generation and editing flows from DALL-E 2/3 to GPT Image models now or keep legacy integrations until shutdown, balancing feature parity, pricing, and migration risk.

Migrate now to GPT Image 1.5 on the Image API, unless image generation is core to a multi-turn conversational flow; keep DALL-E only as a brief bridge to 2026-05-12.

Blockers

Who this is for

Candidates

Migrate now to GPT Image 1.5 on the Image API

As of 2026-03-15, OpenAI says `gpt-image-1.5` is its latest and most advanced image generation model and recommends it for the best experience. The official pricing page lists 1024x1024 image generation at "$0.009" for low quality, "$0.034" for medium, and "$0.133" for high.

When to choose

Best for low-ops + small-team or enterprise + compliance teams where the main workload is single-prompt generation or editing rather than conversational image workflows, and where better instruction following, text rendering, detailed editing, and current-platform support matter more than preserving DALL-E-specific behavior. As of 2026-03-15, this is the clean default migration target if you want off deprecated models without redesigning your product around multi-turn chat.

Tradeoffs

Compared with the Responses API, this is the simpler migration because it stays close to legacy one-shot `images.generate` and `images.edit` patterns, but you give up built-in multi-turn image context and conversational editing flows. Compared with `gpt-image-1-mini`, you get OpenAI's best image experience and stronger quality, but at materially higher per-image cost, so it is a worse fit for bulk or background jobs where price matters more than top-end output.

Cautions

OpenAI says you may need API Organization Verification before using GPT Image models. High input fidelity uses more image input tokens per request, and if you want conversational or iterative editing with image context over multiple turns, OpenAI recommends the Responses API instead of the plain Image API.

Sources

Migrate now to GPT Image via the Responses API

As of 2026-03-15, OpenAI documents the Responses API as the preferred option for conversational, editable image experiences with GPT Image. The image generation tool supports `gpt-image-1.5`, `gpt-image-1`, and `gpt-image-1-mini`, and the guide says Responses adds multi-turn editing and support for image File IDs in context.

When to choose

Best for serverless + low-ops or small-team + real-time style product flows where image generation is part of a conversation, an agent loop, or an iterative editor, and where you want to keep image inputs and outputs in context across turns. Choose this when migration is also a product upgrade, not just a deprecated-model replacement.

Tradeoffs

Compared with the Image API, this gives you multi-turn editing and richer context handling, but the integration is heavier because you must adopt the Responses tool-call shape instead of a near-drop-in model swap. It is also more operationally complex for simple one-shot generation or editing paths, where the extra conversation state and orchestration add migration work without much benefit. Compared with choosing `gpt-image-1-mini` on either API surface, this candidate is about workflow flexibility, not lower cost.

Cautions

OpenAI documents that if you force `action` to `edit` without providing an image in context, the call returns an error. This path also changes the integration shape compared with legacy `images.generate` or `images.edit` flows, so it is a larger migration than a straight model swap.

Sources

Migrate cost-sensitive paths to gpt-image-1-mini

As of 2026-03-15, OpenAI describes `gpt-image-1-mini` as a cost-efficient version of GPT Image and recommends it when image quality is not a priority. The official pricing page lists 1024x1024 image generation at "$0.005" for low quality, "$0.011" for medium, and "$0.036" for high.

When to choose

Best for cost-sensitive + high-scale or serverless + small-team workloads where image generation volume matters more than top-end fidelity, and where you want the GPT Image API surface now without paying `gpt-image-1.5` rates. As of 2026-03-15, this is the pragmatic migration target for bulk or background image jobs that do not need the best text rendering or strongest prompt adherence.

Tradeoffs

This is the cheapest modern migration path, and lower cost is the main reason to choose it. Compared with `gpt-image-1.5`, you are accepting weaker image quality and a lower-priority option when text rendering, prompt adherence, or premium visual output matter. Compared with moving to the Responses API for richer workflows, `gpt-image-1-mini` does not solve conversational product needs by itself; it mainly reduces spend while keeping you on the current GPT Image stack.

Cautions

OpenAI's model page labels `gpt-image-1-mini` as cost-efficient but also marks its speed as "Slowest," so do not assume it is the latency-optimized option. The docs also position it as the lower-priority choice when image quality matters; OpenAI recommends `gpt-image-1.5` for the best experience.

Sources

Keep DALL-E 2 or DALL-E 3 only as a short-lived bridge until shutdown

As of 2026-03-15, this is a containment strategy, not OpenAI's forward path. OpenAI's deprecations page says both `dall-e-2` and `dall-e-3` will be removed on "2026-05-12" and recommends `gpt-image-1` or `gpt-image-1-mini` as replacements.

When to choose

Best only for frozen legacy + low-ops situations where you need a brief bridge while scheduling migration work, especially if you still depend on DALL-E-specific behavior such as DALL-E 2 image variations or an existing DALL-E 2 mask-based flow that you cannot swap immediately. Use this only when the integration is effectively in maintenance mode and you can commit to removal before 2026-05-12.

Tradeoffs

DALL-E 2 still has official documentation for generations, edits, and variations, and OpenAI's image guide calls out lower cost, concurrent requests, and inpainting with a mask. DALL-E 3 remains the higher-quality legacy generation-only option with larger published resolutions than DALL-E 2, but it does not have image edits.

Cautions

Do not start new work on DALL-E 2 or DALL-E 3. OpenAI says DALL-E 3 image edits are not available and recommends GPT Image for editing, and the guide says image variations are available for DALL-E 2 only. Both models have a fixed shutdown date of 2026-05-12, so every week spent on legacy code increases forced-migration risk.

Sources

Facts updated: 2026-03-15
Published: 2026-04-03

Related questions

Try with your AI agent

$ npm install -g pocketlantern
$ pocketlantern init
# Restart Claude Code, Cursor, or your MCP client, then ask:
# "OpenAI Image Stack — when and how should I migrate?"
Missing something? Request coverage