Imagen 4 Is Retired: How to Migrate to Nano Banana 2 and Pro
Short answer: Google shut down the Imagen 4 API family on August 17, 2026 and points every user to the Gemini image models: Nano Banana 2 for everyday work, Nano Banana Pro for the highest quality, Nano Banana 2 Lite for volume. Your prompts move over almost unchanged. What does change: no negative-prompt field, PNG output with a SynthID watermark, one image per request, and reference images plus conversational edits instead of separate edit endpoints.
Model-ID mapping
| Imagen (retired) | Use now | Note |
|---|---|---|
| imagen-4.0-ultra-generate-001 | gemini-3-pro-image (Nano Banana Pro) | Best quality, 4K, up to 14 reference objects / 5 people |
| imagen-4.0-generate-001 | gemini-3.1-flash-image (Nano Banana 2) | Everyday generation and editing, 0.5K–4K |
| imagen-4.0-fast-generate-001 | gemini-3.1-flash-lite-image (Nano Banana 2 Lite) | Cheapest tier for volume ($0.0336 per image) |
| imagen-3.0-capability-001 (edit / inpaint) | Conversational edits on Nano Banana 2 / Pro | Send the image and describe the change; no mask endpoint |
| negativePrompt parameter | Constraints written into the prompt | Not supported on Gemini image models |
| JPEG / WebP output | PNG only, SynthID always embedded | Convert downstream if you need JPEG |
What changes in your prompts
- No negative prompts. Google says so explicitly. Turn "negativePrompt: text, watermark, people" into a sentence: "clean background, no text, no people".
- Write like a creative director. Google's own Nano Banana Pro guidance: no tag soups ("4k, trending on artstation, masterpiece"); describe subject, setting, light, lens and mood in plain sentences.
- Identity comes from references. Nano Banana Pro accepts up to 14 reference objects and keeps up to five people consistent. Attach the canon photo and say what it controls: "use the reference for identity, face and body only — not its pose, background or lighting".
- Edit by conversation. Instead of an inpaint endpoint, send the image back with the change described ("replace the sofa with a linen sectional, keep everything else"). Iterate instead of regenerating from scratch.
- Output is PNG with SynthID. Every image carries the watermark; keep it — the EU AI Act's transparency rules (Article 50) have applied since August 2, 2026.
Prices in September 2026 (Gemini API)
Nano Banana 2: $0.045 (0.5K), $0.067 (1K), $0.101 (2K), $0.151 (4K) per image, half of that in batch mode. Nano Banana 2 Lite: $0.0336 per image. Nano Banana Pro: $0.134 at 1K and 2K, $0.24 at 4K. If you generated hundreds of Imagen Fast images a day, Lite in batch mode is the direct successor; if you used Ultra for hero shots, Pro at 2K is the like-for-like move.
Alternatives if you are re-evaluating anyway
The retirement is a natural moment to test the field. GPT Image 2.5 (September 2026) leads instruction following and text; FLUX.2 leads controllable production pipelines and has open dev weights; Seedream 5.0 Pro leads pixel-level editing; Ideogram 4.0 accepts JSON layouts for posters. See the best AI image generators in 2026 for the full comparison.
Migration checklist
- Replace the model IDs (table above) and remove the negativePrompt parameter.
- Rewrite negatives as constraints inside the prompt; delete quality tag soups.
- Move identity work to reference images with an explicit role sentence.
- Switch downstream code from JPEG/WebP to PNG input; keep SynthID intact.
- Turn on batch mode for non-urgent volume to halve the price.
- Re-run your five most important prompts on Nano Banana 2 and Pro, compare, lock the default.
FAQ
Is Imagen 4 really gone?
Yes. Google retired the Imagen family: the Gemini API models imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001 and imagen-3.0-capability-001 were shut down on August 17, 2026, and Vertex AI customers were asked to migrate earlier. Imagen no longer appears on the Gemini API pricing page.
Which model replaces Imagen 4 Ultra?
Nano Banana Pro (gemini-3-pro-image) for the highest quality, including 4K output and up to 14 reference objects. Nano Banana 2 (gemini-3.1-flash-image) replaces Imagen 4 standard for everyday work, and Nano Banana 2 Lite (gemini-3.1-flash-lite-image) replaces Imagen 4 Fast for volume.
Do Nano Banana models support negative prompts?
No. Google's migration guide states that negative prompts are not supported on Gemini image models. Write exclusions as plain constraints inside the prompt ("clean background, no text, no people").
What else changes technically?
Gemini image models return one image per request, output PNG only, always embed a SynthID watermark, and accept reference images plus conversational follow-up edits instead of a separate edit endpoint. Batch mode halves the price for non-urgent work.
What does it cost now?
On the Gemini API in September 2026: Nano Banana 2 $0.045 (0.5K), $0.067 (1K), $0.101 (2K), $0.151 (4K) per image, half in batch mode; Nano Banana 2 Lite $0.0336; Nano Banana Pro $0.134 at 1K/2K and $0.24 at 4K. Compare with your old Imagen tier before you move volume.
Can I keep my Imagen prompts?
Mostly. Imagen liked descriptive photographic prompts and so does Nano Banana. Remove any negative-prompt field, drop tag soups like "4k, masterpiece", write the brief as a creative director would, and move identity work to reference images.
Need prompts that already speak Nano Banana? GoldenPrompts writes natural-language briefs tuned for Nano Banana Pro, GPT Image 2.5, FLUX.2 and Seedream 5.0 from a few clicks — no negatives, no tag soups. Free to start: 24 hours of everything, no card.