gpt-image-1 sunsets 23 Oct 2026: migrating Bubble bulk catalogs to gpt-image-2 with BIAIG

Why now

OpenAI lists gpt-image-1 shutdown on 2026-10-23 (Deprecations). After that date, calls that still send the gpt-image-1 model id should fail at the API; assets already stored outside OpenAI are not removed by the deprecation itself (Image prompting guide).

If you run production bulk catalogs in Bubble—CSV chunks, async polling, optional batch queues—this is a deadline-driven cutover for your orchestrator. In this editorial thread that orchestrator is BIAIG between Bubble workflows and providers. Scout found no public docs under the name “BIAIG”; the checklist below is grounded in OpenAI’s migration guidance plus common async gateway patterns (e.g. Buble GPT Image 2). CSV-chunk column specifics are not publicly documented—validate against your internal inventory. A closer look: Mixboard offline on 28.09: exporting. Worth reading next: BulkImagen async batches with nano-banana-2-1:. A closer look: Mixboard offline on 28.09: exporting.

Official replacements include gpt-image-2.5-sunburst (quality) and gpt-image-2.5-flare (speed); guides also center gpt-image-2. Confirm which ids BIAIG exposes in production before you regen the full catalog.

---

What breaks on 2026-10-23

Audit every path that still defaults to gpt-image-1:

• BIAIG / backend workflow defaults;
• prompt templates with embedded model fields;
• CSV import mappings;
• retry logic that re-submits the old id.

Rows that already point at durable Bubble/S3/CDN blobs do not require regen for API continuity—only new generations break.

---

Parameters: not a one-line model swap

GPT Image 1 (gpt-image-1)	GPT Image 2 / 2.5
size: 1024x1024, 1024x1536, 1536x1024, auto	auto or custom WIDTHxHEIGHT (16px grid, aspect between 1:3 and 3:1, max edge 3840px)
optional input_fidelity	no input_fidelity—high fidelity is implicit (Image prompting)

If BIAIG (or a gateway such as public openai/gpt-image-2) uses aspect_ratio and resolution (1K/2K/4K) instead of raw OpenAI pixel sizes, build an explicit legacy size → aspect + resolution table and re-check downstream CSV consumers (thumbnails, crops, print layouts).

GPT Image 2.5 adds gpt-image-2.5-flare / gpt-image-2.5-sunburst and extended quality tiers including xhigh/max—that changes cost, latency, and QA bars, not just the model string.

---

Migration checklist (Bubble + BIAIG)

1. Inventory

• Export active prompt templates and reference images used in edit chains.
• Classify jobs: synchronous API vs OpenAI Batch for /v1/images/generations and /v1/images/edits (Batch API: ~50% lower cost vs sync, up to 24h completion, often faster) vs BIAIG async poll (create task → poll until terminal state—see Buble API overview).
• Note where CSV chunks store provider URLs vs durable blobs—Bubble image plugins illustrate URL expiry risk separate from model shutdown. Worth reading next: GPT Image 2.5 at 250. A closer look: BulkImagen SKU templates and POST.

2. Baseline and locked dimensions

Follow OpenAI’s published workflow (Image prompting guide):

1. Lock a reference set of prompts and edit references.
2. First comparison run uses fixed dimensions/format.
3. Post-batch QA: instruction following, text legibility, identity/product geometry, transparency alpha, multi-step edits.

In Bubble, run a small migration QA CSV chunk (tens of rows), not the full catalog.

3. Prompt templates

• Remove input_fidelity dependencies; rewrite prompts that assumed v1 fidelity toggles.
• Expect visual drift on bulk regen even with identical prompt text.
• Version templates: store prompt_template_version and model_id on chunk metadata for partial rollback.

4. Aspect ratio and layout-sensitive SKUs

• Map 1024x1024 → 1:1 + chosen resolution; portrait/landscape legacy sizes → 9:16 / 16:9 (verify on QA—do not assume letterboxing matches v1).
• SKUs with strict downstream crops are prime candidates for mandatory regen after size policy changes.

5. Regenerate vs keep in storage

Action	When
Keep existing blob/URL	Durable storage; CSV metadata still valid; no v1-only edit chain
Regenerate	Missing asset, dead URL, critical text-in-image, PNG alpha, compliance-sensitive product geometry, edit chains tied to v1
QA only	Model changed but pixel drift acceptable for that SKU—product decision

Avoid stopping jobs: use a pending_migration row state until staging confirms the new model path.

6. CSV-chunk continuity

No public spec for your chunk format—minimum contract:

• Idempotent chunks (chunk_id reruns do not double-upload).
• States: queued → provider_pending → stored → qa_pass|qa_fail.
• After cutover, never retry failed rows with model=gpt-image-1 without updating the mapper.

If chunks assume sync completion but migration uses Batch or gateway polling, add timeouts and backpressure to prevent deadlocks.

7. Throughput and verification

• Documented IPM limits on gpt-image-2 (e.g. low tiers) can stall bulk CSV jobs.
• Batch uses a separate rate-limit pool and the documented ~50% batch discount (Batch API).
• Confirm Organization Verification for GPT Image models before peak load (Image generation).

8. Rollout

1. Staging: BIAIG on target gpt-image-2 or 2.5 id + QA chunk.
2. Production: canary slice or new-SKU-only pipeline.
3. Complete cutover before 2026-10-23 with alerts on API failures mentioning deprecated model ids. A closer look: The ~500-row bulk-tools ceiling in.

---

Open questions (resolve before cutover)

• What is BIAIG in your stack (plugin, gateway, direct OpenAI, hybrid) and where is model catalog versioning documented?
• Target gpt-image-2 only (as in Buble public docs) or official 2.5 flare/sunburst, and is 2.5 available through BIAIG today?
• Does “batch discount” mean OpenAI Batch API or BIAIG/gateway credits (Pricing vs per-image gateway credits)?

---

Sources

OpenAI: Deprecations, Image prompting, Batch API, GPT-Image-2, Image generation, Pricing. Gateway/async context: Buble, Bubble plugin pattern. Cross-check: Community deprecation notice.