Google shut down the Imagen 4 API on August 17, 2026. The best AI image generation API for most teams is now Gemini 3.1 Flash Lite Image at $0.0336 per 1K image — 16% cheaper than the Imagen 4 Standard tier it replaces. If you need top-ranked quality, GPT Image 2 leads the Artificial Analysis arena at 1,370 Elo. Google’s own recommended migration path costs 67.5% more.
What exactly shut down on August 17, 2026?
Google turned off three Imagen 4 endpoints yesterday: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001. There is no grace period and no read-only fallback. Google’s own documentation tells developers to move to Gemini 3.1 Flash Image “to avoid service interruptions.”
That is a support answer, not a procurement answer. It ignores what the replacement costs.
Every product that called those endpoints — thumbnail pipelines, e-commerce mockups, ad variant generators — returned an error this week unless someone repointed the client.
The prices you just lost
Imagen 4 was cheap. Fast ran $0.02 per image, Standard $0.04, and Ultra $0.06, per Google’s Imagen documentation and the migration notices that followed it.
Nothing in Google’s current image lineup matches the Standard tier at that price. The company retired its own price leader.
Which is the best AI image generation API right now?
There is no single winner. GPT Image 2 takes raw quality at 1,370 Elo. Gemini 3.1 Flash Lite Image takes value at $0.0336 per 1K image and 1,291 Elo. FLUX.2 [klein] takes the price floor, from $0.014. Your budget picks the model, not your taste.
The Artificial Analysis Image Arena ranks models on blind user votes. Prices below come from each provider’s published rate card.
Price and quality, side by side
| Model | Provider | Arena Elo | Per 1K image | Per 10,000 images |
|---|---|---|---|---|
| GPT Image 2 (high) | OpenAI | 1,370 | $0.165–$0.211 | $1,650–$2,110 |
| Reve 2.1 (v2 endpoint) | Reve | 1,323 | ~$0.20 | ~$2,000 |
| Gemini 3.1 Flash Image | 1,321 | $0.067 | $670 | |
| Gemini 3 Pro Image | 1,297 | $0.134 | $1,340 | |
| Gemini 3.1 Flash Lite Image | 1,291 | $0.0336 | $336 | |
| GPT Image 2 (low) | OpenAI | not separately ranked | $0.005–$0.006 | $50–$60 |
| FLUX.2 [pro] | Black Forest Labs | outside arena top 10 | from $0.03 | from $300 |
| FLUX.2 [klein] 4B | Black Forest Labs | outside arena top 10 | from $0.014 | from $140 |
| Imagen 4 Standard | retired Aug 17, 2026 | $0.04 | $400 |
Read the last row against the fifth. The only Google model that undercuts the Imagen 4 tier you were paying for is Flash Lite — and it is the one Google did not recommend.
Is Google’s recommended migration the right one?
No. Google points Imagen 4 users at Gemini 3.1 Flash Image, which runs $0.067 per 1K image against Imagen 4 Standard’s $0.04. That is a 67.5% increase: $670 versus $400 per 10,000 images. For a team shipping 500,000 images a year, the default migration adds roughly $13,500 in annual spend.
Google’s Gemini API pricing page bills image output by token. Flash Image costs $60 per million output tokens, and a 1K image consumes 1,120 tokens.
The Flash Lite loophole
Gemini 3.1 Flash Lite Image bills at $30 per million output tokens — half of Flash Image. Same 1,120 tokens for a 1K image, so $0.0336 apiece.
It scores 1,291 Elo against Flash Image’s 1,321. You give up 30 Elo points and cut your bill by half.
Batch mode halves it again. Google discounts batch output to $15 per million tokens, which puts a 1K image at roughly $0.0168 — about 58% below the Imagen 4 Standard price you lost.
How much does an AI image generation API cost per 10,000 images?
Between $50 and $2,110, depending entirely on which tier you call. That 42x spread is the single most consequential decision in an image pipeline — larger than any quality gap between the top and bottom of the leaderboard.
Here is the real monthly math at 10,000 1K images:
- FLUX.2 [klein] 4B: from $140 — the cheapest hosted option from a major lab
- Gemini 3.1 Flash Lite Image, batch: ~$168
- FLUX.2 [pro]: from $300
- Gemini 3.1 Flash Lite Image, standard: $336
- Imagen 4 Standard (now dead): $400
- Gemini 3.1 Flash Image: $670
- Gemini 3 Pro Image: $1,340
- GPT Image 2 (high): $1,650–$2,110
GPT Image 2’s low-quality tier at $0.005–$0.006 undercuts everything on this list. It is the right call for internal previews and A/B thumbnails where nobody zooms in.
This is the same cost-per-unit discipline we applied to video in our breakdown of the cheapest AI video APIs before the Sora 2 sunset. Image workloads punish sloppy tier selection harder, because volume is 100x higher.
Which is better for quality, GPT Image 2 or Nano Banana 2?
GPT Image 2, by 49 Elo points — 1,370 versus 1,321 in the text-to-image arena. But you pay $0.165 to $0.211 per high-quality image against $0.067 for Gemini 3.1 Flash Image. That is 2.5x to 3x the price for a roughly 3.7% Elo gain.
OpenAI released GPT Image 2 on June 24, 2026, and it has held the arena’s top slot since. Widescreen and portrait ratios price slightly below square at the medium and high tiers.
The verdict depends on who sees the output. Customer-facing hero images justify GPT Image 2. Bulk catalog art does not.
It is the same trade we mapped when comparing Gemini 3.7 Flash against Claude Sonnet 5 on cost per coding point. Frontier tiers sell small quality deltas at large multiples.
What about image editing, not just generation?
The ranking flips. Reve 2.1 leads the Artificial Analysis editing arena at 1,263 Elo, ahead of GPT Image 2 (high) and MAI-Image-2.5, tied at 1,257. Gemini 3.1 Flash Image sits fifth at 1,249.
Reve’s pricing is bifurcated in a way most buyers miss. The v2 create and edit endpoints run 150 credits, roughly $0.20 per call.
The fast edit endpoint runs 5 credits — about $0.007. That is 28x cheaper than the v2 endpoint from the same vendor holding the top editing score.
FLUX.2 [pro] editing starts at $0.045 per image, a $0.015 premium over its generation price.
Which AI image generation API should you pick for your use case?
Match the model to who consumes the image. Bulk internal assets go to the cheapest tier that clears your bar. Customer-facing work goes to the arena leader. Editing-heavy workflows go to Reve, not to whichever model you already use for generation.
| Use case | Pick | Why | 10,000 images |
|---|---|---|---|
| Direct Imagen 4 replacement | Gemini 3.1 Flash Lite Image | Cheaper than the tier you lost, minimal code change | $336 |
| Highest possible quality | GPT Image 2 (high) | 1,370 Elo, the arena leader | $1,650–$2,110 |
| Bulk catalog / e-commerce | FLUX.2 [klein] 4B | Lowest floor price from a major lab | from $140 |
| Internal previews, thumbnails | GPT Image 2 (low) | $0.005 per image, same API as your high tier | $50–$60 |
| Image editing at scale | Reve 2.1 fast edit | Top editing Elo, ~$0.007 per edit | ~$70 |
| Offline / overnight jobs | Gemini 3.1 Flash Lite, batch | 50% batch discount, no quality change | ~$168 |
| 4K marketing assets | Gemini 3 Pro Image | $0.24 at 4K, cheaper than GPT Image 2 high | $2,400 at 4K |
Frequently asked questions
Can I still call the Imagen 4 API?
No. Google shut down the standard, ultra and fast endpoints on August 17, 2026. Once a model is shut down it is fully off — requests fail rather than degrade.
What is the cheapest AI image generation API in 2026?
GPT Image 2’s low-quality tier at $0.005–$0.006 per image is the cheapest from a top-tier lab. Among models that hold a competitive arena score, FLUX.2 [klein] 4B from $0.014 and Gemini 3.1 Flash Lite Image at $0.0336 lead.
Is Nano Banana 2 the same as Gemini 3.1 Flash Image?
Yes. Nano Banana 2 is the informal name for Gemini 3.1 Flash Image, listed on the arena as Gemini 3.1 Flash Image Preview. Nano Banana Pro refers to Gemini 3 Pro Image.
Why did my Google image bill go up after migrating?
Because Google’s recommended target, Gemini 3.1 Flash Image, costs $0.067 per 1K image versus Imagen 4 Standard’s $0.04 — a 67.5% increase. Switching to Flash Lite Image instead brings you to $0.0336.
Does resolution change the price?
Yes, because Gemini bills images by output token. On Flash Image, 0.5K costs $0.045, 1K costs $0.067, 2K costs $0.101 and 4K costs $0.151. Do not request 4K for assets you display at 800 pixels.
Is GPT Image 1 still safe to build on?
No. GPT Image 1 is scheduled for deprecation on October 23, 2026. Anything new should target GPT Image 2 directly.
Should I self-host an open-weights image model instead?
Only above roughly 100,000 images a month, once GPU amortization beats per-image API pricing. We ran that calculation for language models in the real cost of self-hosting Qwen3.8-Max, and the crossover logic is identical.
The bottom line
Migrate to Gemini 3.1 Flash Lite Image, not Gemini 3.1 Flash Image. It costs $0.0336 per 1K image against Flash Image’s $0.067 and Imagen 4 Standard’s old $0.04. You end up 16% cheaper than before the shutdown, while Google’s recommended path leaves you 67.5% more expensive.
The 30 Elo points you surrender — 1,291 versus 1,321 — are invisible in catalog images, thumbnails and background art. That is where the volume lives.
Reserve GPT Image 2 (high) for the images a paying customer actually looks at. At $0.165 to $0.211 each, it is a marketing line item, not an infrastructure one.
If your workload is edit-heavy, move it to Reve 2.1’s fast edit endpoint at roughly $0.007 per call. It holds the top editing score at 1,263 Elo and costs a fraction of the v2 endpoint the same vendor sells.
One rule beats all of this: bill by tier, not by brand. The gap between a provider’s cheapest and priciest tier is 42x. The gap between providers at the same tier is rounding error. Vendors keep proving that — as we saw when DeepSeek raised prices up to 1,100% overnight. Price your pipeline on the tier, and the brand stops mattering.
Sources
- Google AI for Developers — Imagen 4 deprecation and shutdown notice
- Google Gemini API — image generation pricing
- Artificial Analysis — Text-to-Image Arena leaderboard
- Artificial Analysis — Image Editing Arena leaderboard
- Black Forest Labs — FLUX API pricing
- OpenRouter — GPT Image 2 specifications and release date
Leave a Reply