Imagen 4 vs Ultra: shutdown dates, old prices & migration
Quick answer
Imagen 4, Imagen 4 Fast and Imagen 4 Ultra are no longer good choices for a new integration. Google Cloud's model documentation lists their generation endpoints as discontinued June 30, 2026, and the Gemini API deprecation schedule lists the corresponding endpoints as shut down August 17, 2026. Google's pricing page still shows the old $0.02, $0.04 and $0.06 per-image rates, but a retained price row does not prove current endpoint access.
A quiet glass observatory above a cloud layer at blue hour, warm interior lights, one telescope, cinematic realistic lighting, balanced composition, fine architectural detail, no people, text, or logos.
Generate a current comparison set
Test a current image model on the same prompt, aspect ratio and acceptance criteria before migrating production volume.
Try AI Image GeneratorImagen 4 vs Ultra at a glance
Route | Documented lifecycle | Former generation price | What to do now |
|---|---|---|---|
Imagen 4 Fast | Google Cloud endpoint discontinued June 30, 2026; Gemini API endpoint shut down August 17, 2026 | $0.02 per image | Do not start a new integration; choose a current image model |
Imagen 4 | Google Cloud endpoint discontinued June 30, 2026; Gemini API endpoint shut down August 17, 2026 | $0.04 per image | Migrate existing calls and re-baseline output quality |
Imagen 4 Ultra | Google Cloud endpoint discontinued June 30, 2026; Gemini API endpoint shut down August 17, 2026 | $0.06 per image | Do not treat the higher archived price as evidence of current access or quality |
Lifecycle and prices were checked September 13, 2026 against Google's Imagen 4 model page, pricing page and Gemini API deprecation schedule. The former prices are included to interpret older budgets and articles, not as a purchasing recommendation.
What was the difference between Imagen 4 and Ultra?
Standard Imagen 4 and Ultra accepted text and returned images. Both documented up to four output images per prompt, the same five aspect ratios, and resolutions up to 2816 by 1536 or 1536 by 2816. Ultra cost 50% more per generated image: $0.06 instead of $0.04.
That price difference never established a universal quality winner. A model is worth more only when it raises the acceptance rate or reduces correction time on a defined job. Google's documentation also listed different request quotas: 75 regional online prediction requests per minute for standard Imagen 4 and 30 for Ultra.
What should an existing Imagen 4 user do?
Inventory every endpoint. Record the exact model ID, platform, region, output dimensions and retry behavior. Do not infer access from an old pricing table.
Choose a current replacement. Google's Gemini API deprecation page points the former Imagen 4 routes to gemini-3.1-flash-image. Confirm the current catalog for the platform where the application runs.
Re-run representative prompts. Keep prompts, aspect ratios, safety settings and sample counts fixed. Compare every output and record manual corrections.
Recalculate cost per accepted image. Include current input and output charges, rejected generations, upscales, storage and editing time.
Update monitoring and documentation. Record the new model ID and migration date so a later lifecycle change is visible.
How to compare the replacement
Use a prompt that exposes the task's real constraints. For product imagery, require the exact object, room for copy, correct colors and believable lighting. For illustration, define the subject, composition, style and forbidden elements. Generate the same number of candidates from each route and score all of them.
Count a candidate as accepted only if it can be published after the normal review. If 100 generations cost $4 and 20 are accepted, generation cost per accepted image is $0.20 before editing. If another route costs $6 but yields 40 accepted images, it is $0.15. These numbers explain the method; they are not measured Imagen performance.
Current image generation and editing are separate choices
Google's current image-generation documentation covers active Gemini image routes. Check it at implementation time because model IDs and preview status can change. The Imagen 4 pricing and API guide preserves more integration and historical cost context, while our Nano Banana Pro review covers a current Google image-model decision.
For a browser workflow, use Magic Hour's AI Image Generator to create a new still. Use the AI Image Editor when the source image must remain recognizable and only a controlled change is needed. Generation and editing should be evaluated separately.
Frequently asked questions
Do not plan a new integration around it. Google's Cloud model page lists the Imagen 4, Fast and Ultra generation endpoints as discontinued June 30, 2026. The Gemini API deprecation schedule lists those model IDs as shut down August 17, 2026.
A pricing page can retain historical or product-family rows after an endpoint lifecycle changes. Verify the model page, endpoint catalog and account access together; pricing alone is not availability evidence.
Ultra cost more and had a lower documented request quota, but those facts do not prove better results for every prompt. A fair conclusion requires the same prompts, settings, sample count and acceptance criteria.
Google's current deprecation schedule points the former Imagen 4, Fast and Ultra routes to gemini-3.1-flash-image. Recheck the current model catalog before implementation because replacements and lifecycle dates can change.
No. Use them only to interpret an older estimate. Build a new budget from the exact current endpoint, platform, output size and retry behavior, then calculate cost per accepted image.
Runbo Li is the Co-founder and CEO of Magic Hour, where he builds AI video and image tools for content creation. He is a Y Combinator W24 founder and former Data Scientist at Meta, where he worked on 0-1 consumer social products in New Product Experimentation. He writes about AI video generation, AI image creation, creative workflows, and creator tools.
