OpenAI gpt-image-2 Migration Guide for gpt-image-1, gpt-image-1.5, and chatgpt-image-latest APIs

On June 2, 2026, OpenAI announced the gpt-image-1-mini deprecation — along with gpt-image-1.5 and chatgpt-image-latest — effective December 1, 2026. If your routing config or async media pipeline still references any of these model IDs, you have six months to migrate to gpt-image-2. This post covers what changed, what the breaking points are, and what operators should action before the shutdown date.

What OpenAI deprecated and when

The June 2 deprecation notice covers three image-generation model IDs:

A fourth model — gpt-image-1 (base variant) — was already scheduled for shutdown on October 23, 2026, in OpenAI's April 22 notice. Together, every gpt-image-1 variant is gone by Q4 2026, and the only supported OpenAI image-generation API is gpt-image-2.

Why this matters for routing and async pipelines

Unlike text model swaps (where output format is stable across generations), migrating image models touches several routing-layer concerns:

1. Model alias resolution in routing configs

If your gateway or router hard-codes gpt-image-1-mini as a model alias or fallback target, calls after December 1 return an error — a silent failure if no alerting covers model-not-found errors at the provider level.

2. Pricing and quota unit changes

gpt-image-2 uses a different pricing structure than its predecessors: output is billed per image at tiered resolution pricing, not by token count. Any cost-model or budget gate in your routing policy that uses per-token cost estimates for image calls will produce incorrect budget projections after migration.

3. Async job pattern compatibility

Teams running image generation through async job queues — common for batch rendering, product catalog pipelines, or content automation — should verify that their queue processors handle gpt-image-2's response format correctly. TheRouter's async media platform routes this model through the unified v2 job interface; see TheRouter v2.0 Unified Async Media Platform for the dispatch and polling contract.

4. chatgpt-image-latest alias risk

chatgpt-image-latest is an alias, not a pinned snapshot. Operators who added it as a convenience pointer may not realize it's being fully retired rather than redirected. Unlike most latest-alias deprecations, there is no automatic forward-pointing to gpt-image-2. The alias will simply fail after December 1.

How to migrate gpt-image-1-mini to gpt-image-2 in routing policy

Follow this checklist to complete the migration before December 1, 2026:

1. Audit model references. Search your routing config, SDK call sites, and any hardcoded model strings for gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest, and gpt-image-1. The last one has an earlier shutdown (October 23).

2. Update fallback chains. If gpt-image-1-mini appears as a cost-optimized fallback in your chain, remove it and point directly to gpt-image-2. Running on sunset models in a fallback position delays migration and reduces observability.

3. Re-validate budget gates. gpt-image-2 pricing tiers (low / medium / high quality × resolution) differ from the previous standard / HD structure. If your routing policy gates on per-request cost, update the cost estimate used in routing decisions.

4. Test response format parity. gpt-image-2 returns base64-encoded image data in the same Images API shape, but verify that downstream consumers handle any quality-tier or size differences your pipeline depends on.

5. Set migration deadline alerts. Add a gate at October 1, 2026 (two months before shutdown) to force a review if either model ID still appears in active routing paths.

6. Consider a non-OpenAI fallback. For teams evaluating cost-optimized alternatives at scale, test a provider-agnostic fallback (e.g., Qwen-Image-2.0 or Imagen 3 via Vertex) as a cold standby while completing the primary migration.

gpt-image-2 capability delta worth knowing

Before migrating, understand what gpt-image-2 adds over the deprecated models:

• Inpainting and outpainting via the edits endpoint — not available in the mini and 1.5 variants
• Higher effective resolution at medium quality than gpt-image-1.5 at HD
• Improved instruction following on complex prompts with multiple objects
• C2PA provenance metadata embedded in output (relevant for governance pipelines)

The improved capability set means the migration is generally an upgrade, not a lateral move — but the pricing model change requires operator awareness.

Timeline summary

The December 1 shutdown gives you roughly 26 weeks from today. Given typical enterprise change-freeze windows in Q3 and Q4, the effective working window is shorter. Start now.

What TheRouter users should do

If you use TheRouter to route image generation requests, update your model list to replace gpt-image-1-mini and gpt-image-1.5 with gpt-image-2 in any routing rules or provider configs. The async media resource delivery docs cover the full job lifecycle for image generation through TheRouter, including how to handle provider-level errors that will replace successful responses after the shutdown date.