Model-lifecycle noise is almost always about chat models. Meanwhile Google's generative media line — image, video, music — has been quietly running the same deprecation machinery for a year, and one of its deadlines lands on Monday, August 17, 2026.
Every number and quote below was read this morning off Google's own pages, not off a tracker:
ai.google.dev/gemini-api/docs/pricing-
ai.google.dev/gemini-api/docs/deprecations(page footer: Last updated 2026-08-03 UTC)
The deadline
Three model ids go dark on August 17:
imagen-4.0-generate-001
imagen-4.0-ultra-generate-001
imagen-4.0-fast-generate-001
They were released June 24, 2025. That's it — a fourteen-month production life for what the pricing page still describes as "Our latest image generation model, with significantly better text rendering and better overall image quality."
Where to go next has two official answers
The pricing page carries this warning, verbatim:
Warning: Imagen 4 models (
imagen-4.0-generate-001,imagen-4.0-ultra-generate-001,imagen-4.0-fast-generate-001) are deprecated and will be shut down on August 17, 2026; migrate to Gemini 2.5 Flash Image to avoid service disruption.
The deprecations table, for the same three ids, puts something else in its Recommended replacement column:
| Model | Release date | Shutdown date | Recommended replacement |
|---|---|---|---|
imagen-4.0-generate-001 |
June 24, 2025 | August 17, 2026 | gemini-3.1-flash-image |
imagen-4.0-ultra-generate-001 |
June 24, 2025 | August 17, 2026 | gemini-3.1-flash-image |
imagen-4.0-fast-generate-001 |
June 24, 2025 | August 17, 2026 | gemini-3.1-flash-image |
Two first-party surfaces, two different destinations. And it gets worse if you follow the banner, because Gemini 2.5 Flash Image is itself deprecated — it's in the same table with a shutdown of October 2, 2026, six weeks after the model it's being offered as an escape from.
Follow the table's pointer for that model and you get this chain:
gemini-2.5-flash-image shutdown Oct 2, 2026
└─ recommended: gemini-3.1-flash-image-preview
shutdown Jun 25, 2026 ← already passed
└─ recommended: gemini-3.1-flash-image
no shutdown date announced
The middle hop is a model whose own shutdown date was 47 days ago. Nothing here is malicious — it's an ordinary docs-consistency problem across three tables that each move on their own schedule. But if you're writing the migration, you need to know that the pointer requires two hops and that one of the hops is dead.
Worth noting what the pricing page does not say: in the Gemini 2.5 Flash Image section there is no deprecation warning at all. It reads as a normal, buyable model. A reader who follows the banner lands on a page that looks clean.
It isn't just an id swap — the billing unit changes
Imagen 4 bills per image, flat, paid tier only (there is no free tier):
| Price per image | |
|---|---|
| Imagen 4 Fast | $0.02 |
| Imagen 4 Standard | $0.04 |
| Imagen 4 Ultra | $0.06 |
The banner's target, Gemini 2.5 Flash Image, also quotes a flat per-image output price: $0.039 per image on the Standard tier.
The table's target, Gemini 3.1 Flash Image, bills per token — image output at $60.00 per 1M tokens — which Google converts on the page to a resolution-tiered rate:
| Resolution | Price per image |
|---|---|
| 0.5K | $0.045 |
| 1K | $0.067 |
| 2K | $0.101 |
| 4K | $0.151 |
So imagen-4.0-generate-001 at $0.04 costs the same whatever you render; its replacement costs $0.045 at 0.5K and $0.151 at 4K — a shade under 3.8x — and image size becomes a line item in your cost model where it previously wasn't one. That's the part that breaks a spreadsheet, not the model id.
A Google shutdown date is a floor, not a schedule
This is the bit worth internalizing, and Google states it plainly on the deprecations page:
Deprecation announcements are made on the Release notes page, and the announced earliest shutdown dates are tracked on this page. Already-shutdown models are indicated with gray backgrounds.
Now look at the Veo rows. veo-3.0-generate-001, veo-3.0-fast-generate-001 and veo-2.0-generate-001 all carry a shutdown date of June 30, 2026 — 42 days ago as I write this. They are not gray. The pricing page still sells them, with live per-second rates and a warning in the future tense: "will be shut down on June 30, 2026."
So the date is a lower bound on when the endpoint disappears, not a prediction of when it will. Both failure modes are real: build a migration that assumes the date holds and you may cut over needlessly; build one that assumes it slips and you find out the hard way. The only safe read is "not before this date."
The stable video model is the deprecated one
Also from the pricing page, about Veo 3 — the deprecated line:
Our stable video generation model, available to developers on the paid tier of the Gemini API.
And about Veo 3.1, the model the deprecations table names as its replacement:
Preview models may change before becoming stable and have more restrictive rate limits.
Every Veo model in the Gemini API is now either deprecated or preview. The table hedges accordingly — the replacement cell reads veo-3.1-generate-preview "or the GA models on the Gemini Enterprise Agent Platform", which is a different product with a different surface. If you need a generally-available video model inside the Gemini API specifically, as of today the pricing page does not offer you one.
Google's generative-media lifecycle, in one table
Every model in Google's Imagen, Veo, Lyria and image-generation sections that carries a shutdown date, oldest deadline first:
| Model | Shutdown | Recommended replacement |
|---|---|---|
imagen-3.0-generate-002 |
Nov 10, 2025 | imagen-4.0-generate-001 |
veo-3.0-generate-preview |
Nov 12, 2025 | veo-3.1-generate-preview |
veo-3.0-fast-generate-preview |
Nov 12, 2025 | veo-3.1-fast-generate-preview |
gemini-2.0-flash-preview-image-generation |
Nov 14, 2025 | gemini-2.5-flash-image |
gemini-2.5-flash-image-preview |
Jan 15, 2026 | gemini-2.5-flash-image |
imagen-4.0-generate-preview-06-06 |
Feb 17, 2026 | imagen-4.0-generate-001 |
imagen-4.0-ultra-generate-preview-06-06 |
Feb 17, 2026 | imagen-4.0-ultra-generate-001 |
gemini-3.1-flash-image-preview |
Jun 25, 2026 | gemini-3.1-flash-image |
gemini-3-pro-image-preview |
Jun 25, 2026 | gemini-3-pro-image |
veo-3.0-generate-001 |
Jun 30, 2026 | veo-3.1-generate-preview |
veo-3.0-fast-generate-001 |
Jun 30, 2026 | veo-3.1-fast-generate-preview |
veo-2.0-generate-001 |
Jun 30, 2026 | veo-3.1-generate-preview |
imagen-4.0-generate-001 |
Aug 17, 2026 | gemini-3.1-flash-image |
imagen-4.0-ultra-generate-001 |
Aug 17, 2026 | gemini-3.1-flash-image |
imagen-4.0-fast-generate-001 |
Aug 17, 2026 | gemini-3.1-flash-image |
gemini-2.5-flash-image |
Oct 2, 2026 | gemini-3.1-flash-image-preview |
The other models in those same sections read "No shutdown date announced": gemini-3.1-flash-image, gemini-3-pro-image, veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview, lyria-3-clip-preview, lyria-3-pro-preview, lyria-realtime-exp.
Two patterns fall out of the column of dates. Preview lifetimes are short and getting shorter — gemini-3.1-flash-image-preview ran from Feb 26 to Jun 25, 2026, about four months, where the Imagen 4 previews got roughly eight. And a lineage gets re-pointed as it goes: Imagen 3 pointed at Imagen 4, which now points out of Imagen entirely and into the Gemini image line.
Why you probably haven't seen this tracked
Because these models don't fit the schema everyone uses. They have no context window and no per-token price — they bill per image or per second of video — so a per-token price tracker either drops them or shows a row of blanks.
That includes ours, partly. AI Model Watch carries all 25 of Google's generative-media models with their lifecycle fields, but our price fields are per-token, so on these rows they're null and the real rate ($0.04 per image, $0.40 per second) lives in the row's notes with the source URL it was read from. A null there means "this model isn't priced in tokens," not "we couldn't find it" — but it's a genuine limitation of the schema and worth saying out loud rather than papering over.
The lifecycle half is structured and free:
curl -s https://aimodelwatch.dev/api/models.json \
| jq -r '.models[]
| select(.provider=="Google")
| select(.modality | any(test("image-out|video-out|audio-out")))
| [(.retires_on // "none"), .id, .status, (.replacement // "-")] | @tsv' \
| sort
That returns all 25 rows, dated deadlines first. (There's also /api/deprecations.json, which is the lifecycle-only projection — it's narrower, and it has no modality field, so filter that one on .provider and .id instead.)
No key, no signup, CORS open, and every row carries the source_url it was read from so you can check it against Google yourself. Which you should — that's the whole point of the source_url being there.
If you're on Imagen 4 right now
- You have until Monday, August 17. Test against the table's target,
gemini-3.1-flash-image, not the banner's — the banner's target has an October date on it. - Re-price before you cut over. Flat per-image becomes resolution-tiered per-token, and 4K lands near 3.8x what Imagen 4 Standard cost you.
- If you're on Veo instead, you have more time than the page implies, and less certainty. June 30 came and went; nothing is gray yet.
The general version, which is the reason I keep a catalog of this at all: a model shutdown is a date on someone else's calendar, and nothing in your build system is watching it.
Top comments (0)