OpenAI Image API Consolidation: gpt-image-1 Sunsets Oct 23, Three More Dec 1 — Migrate to gpt-image-2

OpenAI is consolidating its entire image generation API into a single model: gpt-image-2. Four image API routes are being retired across two deadlines this fall, per OpenAI’s Deprecations page.

• October 23, 2026: gpt-image-1 sunsets
• December 1, 2026: gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest sunset

Calls to retired routes return errors after their deadline, per the same Deprecations page.

This is a research-based summary. We have not tested any of these endpoints directly.

The migration treadmill

To understand why this announcement stings, trace the timeline:

May 12, 2026 — DALL-E 2 and DALL-E 3 retire, as OpenAI announced on November 14, 2025. OpenAI’s documented migration path pointed developers at gpt-image-2, gpt-image-1, or gpt-image-1-mini as replacements.

April 21, 2026 — gpt-image-2 releases. OpenAI’s next-generation image model becomes available in the API and Codex the same day (OpenAI, “Introducing ChatGPT Images 2.0”; OpenAI Developer Community announcement).

June 2, 2026 — OpenAI notifies developers: gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest will retire December 1, 2026 (OpenAI Deprecations page; community deprecation notice). gpt-image-1 gets a separate October 23 sunset, also listed on the Deprecations page.

The result: developers who picked gpt-image-1-mini — one of OpenAI’s three listed replacements — when migrating off DALL-E 3 in May 2026 now need to migrate again before December 1 — a second forced migration within roughly six months.

Sunset timeline at a glance

What gpt-image-2 adds

Released April 21, 2026, gpt-image-2 is architecturally different from the gpt-image-1 family (OpenAI, “Introducing ChatGPT Images 2.0”). The key additions:

Agentic reasoning before generation. OpenAI describes the model as a “visual thought partner” that can “research, reason, transform source materials, and generate polished visuals end-to-end” before producing pixels (OpenAI launch post; OpenAI Developer Community announcement). This is reported to materially improve complex scenes, spatial relationships, and prompts requiring factual accuracy (The Decoder coverage).

Improved text rendering. OpenAI’s own API guide says text rendering is “significantly improved” over prior models, though it “can still struggle with precise text placement and clarity” — so treat “near-perfect” claims from third-party coverage with some skepticism. The launch post demonstrates rendering across multiple non-Latin scripts, including Japanese, Korean, Chinese, Arabic, Devanagari (Hindi), and Bengali (OpenAI launch post); independent write-ups report noticeably fewer artifacts on dense text and UI mockups than earlier models (VentureBeat).

Web search integration. OpenAI’s launch post references “thinking mode search capabilities,” and reporting on the launch describes the model retrieving current information (e.g., a brand’s current logo or a venue’s appearance) before generating, to work around its training cutoff (OpenAI launch post; The Decoder coverage).

Higher resolution output. Support extends up to 3840×2160 (4K); custom dimensions must have both edges as multiples of 16px, a maximum edge of 3840px, and total pixels between 655,360 and 8,294,400, per OpenAI’s image generation guide. Note OpenAI’s own launch materials describe resolution “up to 2K” as the headline spec, with 4K available as a documented API option.

Request shape change

Contrary to some third-party migration write-ups, gpt-image-2 does not drop the quality parameter — per OpenAI’s image generation guide, it accepts the same quality values (low, medium, high, auto) as the gpt-image-1.x family, used alongside size:

# gpt-image-1-mini (unchanged parameter shape)
response = client.images.generate(
    model="gpt-image-1-mini",
    prompt="a product photo of a ceramic mug on a white background",
    quality="medium",  # low | medium | high
    size="1024x1024",
)
image_b64 = response.data[0].b64_json

# gpt-image-2 — same parameter names, new size ceiling
response = client.images.generate(
    model="gpt-image-2",
    prompt="a product photo of a ceramic mug on a white background",
    quality="medium",   # low | medium | high | auto
    size="1024x1024",   # 1K square
    # size="2048x2048"  # 2K square
    # size="3840x2160"  # 4K landscape
)
image_b64 = response.data[0].b64_json

The response format (data[0].b64_json) is unchanged and the model string changes, but the real shape change is on the size side: gpt-image-2 adds 2K and 4K presets (2048x2048, 3840x2160, 2160x3840) beyond the gpt-image-1.x family’s 1K-class sizes, and OpenAI flags anything above 2560x1440 total pixels (roughly 2K) as “experimental” (image generation guide). gpt-image-2 also drops support for background: "transparent", which the gpt-image-1.x family supports — a real regression to check for if your pipeline generates transparent PNGs (image generation guide).

For image editing (client.images.edit()), the same model-substitution pattern applies, but gpt-image-2 always processes reference images at high fidelity and does not let you configure fidelity down, per the same guide.

Cost impact by tier

All of these models — old and new — are billed on token-based pricing, not a flat per-image rate. OpenAI’s image generation guide gives these per-image cost estimates at standard 1024×1024:

For mini users: migration to gpt-image-2 at equivalent low-output size is roughly a 20% price increase ($0.005 → ~$0.006). Significant at volume but not catastrophic.

For gpt-image-1.5 users at high quality: migration to gpt-image-2 high is a roughly 59% price increase ($0.133 → ~$0.211). Budget for this before the deadline.

Cost mitigation: OpenAI’s Batch API already applies to image generation models today, halves token rates, and runs on a 24-hour completion window (per the Batch API guide and pricing page). For high-volume generation pipelines that aren’t latency-sensitive, batching drops effective per-image cost roughly in half. Per the same pricing page, cached text inputs for gpt-image-2 fall from $5.00 to $1.25 per million tokens when the same prompt is reused at scale.

Edit requests cost more than generation-only requests: gpt-image-2 always processes the reference image at high fidelity (this isn’t configurable down, per the image generation guide), so you’re paying image input tokens for the reference image on top of output tokens. Factor this into iterative workflows.

Migration checklist

Before October 23 (gpt-image-1 users):

• Identify all model="gpt-image-1" call sites in production
• Test gpt-image-2 prompts against your existing prompt library
• Swap the model string, keep quality and size (both still apply to gpt-image-2), and check whether you want the new 2K/4K size presets
• Update cost monitoring for the new per-size, per-quality token rates (pricing page)
• If you rely on background: "transparent", note gpt-image-2 does not support it (image generation guide)
• Deploy to staging, verify output quality is acceptable for your use case

Before December 1 (gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest users):

• Same steps as above for each affected route
• Recalculate unit economics — especially for gpt-image-1.5 high-quality users facing a ~59% per-image increase
• Evaluate Batch API for non-latency-sensitive generation to halve costs
• Re-check any logic keyed to quality string values — the parameter still exists on gpt-image-2, but the price-per-tier and size ceiling changed

General:

• Update error handling — the retired model routes return errors, not degraded results, per OpenAI’s Deprecations page
• Do not alias chatgpt-image-latest and expect continuity — this route is retiring too
• Re-test your existing prompt library against gpt-image-2 rather than assuming a like-for-like swap; independent coverage reports fewer artifacts on dense text and better instruction-following, but “quality” is judged per use case (VentureBeat)

What to watch

Pricing adjustments: OpenAI has cut prices on other models as usage scaled and competitive pressure increased — for example, cutting GPT-5.6 Terra and Luna pricing in July 2026. It is plausible — not confirmed — that gpt-image-2 pricing drops after the consolidation completes.

gpt-image-1-mini replacement: Whether OpenAI will offer a budget-tier image model after December 1 has been raised by developers on OpenAI’s community forum, but no OpenAI staff announcement has been made as of this writing. For now, gpt-image-2 at low quality is the closest analog.

Resources

• OpenAI Deprecations page
• OpenAI Image Generation guide
• OpenAI Community — deprecation notice thread
• AI Weekly — OpenAI Deprecates Three Image APIs
• Unifically — GPT Image 1 vs 1.5 vs 2 comparison