Google Imagen 4 (May 2025) is a typography-strong text-to-image family in three tiers — Fast ($0.02), Standard ($0.04), Ultra ($0.06) — with 1K/2K output and always-on SynthID. It is also a sunset line: deprecated on Vertex AI March 24, 2026, and shut down on the Gemini API August 17, 2026, with Google routing image work to the Gemini-native Nano Banana models. Imagen 4 still appears in invideo's image roster today.
Updated August 2026
Google Imagen 4 is a strong text-to-image model with best-in-class typography — and it is a sunset product. Google deprecated every Imagen 3 and 4 endpoint on Vertex AI on March 24, 2026, and shuts Imagen 4 down on the Gemini API on August 17, 2026, routing image work to its Gemini-native models (the Nano Banana line). Most "Imagen 4 guides" published in 2026 never mention that you soon can't call it — this one is built around that fact, down to a marketing-versus-docs contradiction in the specs and a price point that dies with the model.
Is Google Imagen shutting down?
Yes, on a published schedule. The dated record, from Google's own release notes and API docs (as of August 2026):
| Date | Event |
|---|---|
| Oct 2, 2025 | Gemini 2.5 Flash Image ("Nano Banana") reaches GA, positioned as the Imagen replacement |
| Dec 31, 2025 | Imagen subject/style tuning removed — the first migration signal |
| Mar 24, 2026 | All Imagen 3.0/4.0 generation endpoints deprecated on Vertex AI; Google's stated migration target is gemini-2.5-flash-image, with a June 30, 2026 migration deadline |
| Aug 17, 2026 | Imagen 4 shut down on the Gemini API (Imagen 3 already shut down) |
Sources: Vertex AI release notes, Gemini API Imagen docs. If you are building anything new on Imagen model IDs today, you are building on borrowed time.
How did Imagen get from 3 to 4?
Imagen 3 (August 2024) brought better detail and lighting over Imagen 2 and opened to all Vertex AI users without an allowlist in December 2024, alongside GA "capability" models for editing — inpainting, outpainting, subject/style customization, product editing. The January 2025 imagen-3.0-generate-002 refresh added an always-on LLM prompt rewriter (Vertex AI release notes).
Imagen 4 was announced May 20–21, 2025 at Google I/O and on Vertex, in three tiers: Generate (standard), Fast, and Ultra (maximum prompt alignment), with 2K output. Its headline advance was typography — accurate spelling, longer strings, and layouts for comics, packaging, and posters — with Google claiming the Fast tier ran "up to 10× faster than our previous model" (Google Cloud launch blog, May 2025). It reached the Gemini API in paid preview on June 24, 2025 and GA on Vertex on August 14, 2025. Notably, Imagen 4 never got its own editing endpoints — editing stayed on the Imagen 3 capability models, which in hindsight was an early sign the line was winding down.
What are Imagen 4's documented specs?
| Spec | Imagen 4 (all tiers) |
|---|---|
| Model IDs | imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001 |
| Mode | Text-to-image only on the Gemini API (no conversational editing, no reference images) |
| Resolution | 1K default; 2K on Standard and Ultra |
| Aspect ratios | 1:1, 3:4, 4:3, 9:16, 16:9 — five, fixed |
| Prompt limit | Up to 480 tokens, 1–4 images per request |
| Prompt language | English only per the API docs (see below) |
| Watermark | SynthID invisible watermark on every output, always on |
| People controls | dont_allow / allow_adult / allow_all |
Source: Gemini API Imagen docs, as of August 2026.
The multilingual contradiction: Google's launch marketing advertised "multilingual prompt support," but the Gemini API documentation restricts Imagen prompts to English only. The marketing claim likely referred to limited language coverage on Vertex, but as written, the two official sources conflict — if you prompt Imagen in another language, you are outside the documented envelope.
How does Imagen 4 rank in 2026?
Mid-pack, with one genuine oddity. On the Arena.ai text-to-image leaderboard (July 10, 2026), Imagen 4 Ultra sits #27 at 1148 Elo and Imagen 4 Standard #34 at 1129 — far behind the #1 model at 1385, and behind Google's own Nano Banana family. On Artificial Analysis (August 2026), the strange part appears: Imagen 3 v002 scores 1108, above Imagen 4 Standard at 1101. In blind human preference, the "4" branding did not uniformly beat 3 — only the Ultra tier (1172) clearly did. Google's own claim of winning GenAI-Bench human evals against "leading text-to-image models" (DeepMind Imagen page) named no competitors and predates these rankings.
How much does Imagen 4 cost?
| Model | Price per image |
|---|---|
| Imagen 4 Fast | $0.02 |
| Imagen 4 Standard | $0.04 |
| Imagen 4 Ultra | $0.06 |
Source: Gemini API pricing, as of August 2026. Imagen 4 Fast at $0.02 is the cheapest first-party image generation Google has ever offered — undercutting even Nano Banana at $0.039 — and that price point dies with the model on August 17, 2026. Imagen 3 pricing is no longer listed at all.
What replaces Imagen? The Nano Banana line
Google's official framing was a division of labor: Imagen as the specialized text-to-image model with the best quality/latency/cost trade-off for pure generation, and the Gemini-native image models — the Nano Banana line — for flexible multimodal creation and editing, with conversational edits, up to 14 reference images, and world knowledge (Vertex AI image overview). In practice, Google picked a winner: every Imagen endpoint's migration path points at gemini-2.5-flash-image and its successors, which now outrank Imagen 4 on every leaderboard. If you are choosing a Google image model today, the realistic decision is between Nano Banana generations, not Imagen tiers — see the Nano Banana 2 hub on invideo for the current line.
What is Imagen 4 still good for — and how do you migrate?
While it remains callable, Imagen 4's typography is its lasting strength: exact display text in quotes renders reliably on posters, packaging mockups, and comic panels — Google's flagship demos were Art-Deco travel posters with verbatim taglines (Google Cloud launch blog). Prompt it Google's documented way — subject + context + style, photography descriptors, exact strings quoted — and that prompt library transfers directly to Nano Banana models, which follow the same conventions and add editing on top. The honest migration note: nothing first-party replaces Fast's $0.02 price; Nano Banana at $0.039 is the closest successor. Practical AI image work — thumbnails, product stills, poster layouts — should treat Imagen 4 as a fallback and Nano Banana as the default.
Imagen questions, answered
Is Google Imagen shutting down?
Yes. All Imagen 3/4 endpoints were deprecated on Vertex AI on March 24, 2026, and Imagen 4 shuts down on the Gemini API on August 17, 2026, per Google's release notes and API docs. Imagen 3 is already gone from the Gemini API.
What replaces Imagen 4?
Google's designated migration target is gemini-2.5-flash-image and the wider Gemini-native line (Nano Banana, Nano Banana 2, Nano Banana Pro), which adds conversational editing and reference images and now outranks Imagen 4 on public leaderboards.
Can I use Imagen 4 images commercially?
Generated images are usable commercially under Google's terms, and Google Cloud has offered generative-AI indemnification covering Vertex image services — but confirm Imagen 4's presence on the current indemnified services list yourself; coverage lists change, and a deprecated model's status deserves a fresh check.
Do Imagen images have a watermark?
Yes — every Imagen output carries Google's SynthID invisible watermark. It is always on and not user-removable (Gemini API docs).
What resolutions and aspect ratios does Imagen 4 support?
1K by default, 2K on Standard and Ultra, across five aspect ratios: 1:1, 3:4, 4:3, 9:16, and 16:9. Prompts are capped at 480 tokens and documented as English-only on the Gemini API.
Why does Imagen 3 sometimes rank above Imagen 4?
On Artificial Analysis' arena (August 2026), Imagen 3 v002 holds 1108 Elo versus Imagen 4 Standard's 1101 — blind voters slightly preferred the older model at the standard tier. Only Imagen 4 Ultra (1172) decisively beat Imagen 3.
Can you still use Imagen 4 today?
Yes — Imagen 4 appears in the image tab of the invideo agent's model roster today, so you can still put its typography to work — or A/B it against Nano Banana Pro and the platform's other image models — without a Google Cloud account. But August 17, 2026 is Google's API shutdown, and once Google retires the model upstream no platform can serve it, so treat Imagen workflows as short-term and let the agent's routing move you to the successors. The Imagen hub on invideo tracks the family's current status inside the platform.
Version history: Imagen 3 Aug 2024 (open Vertex access Dec 10, 2024; 002 refresh with prompt rewriter Jan 29, 2025) → Imagen 4 announced May 20–21, 2025 (Generate/Fast/Ultra, 2K) → Gemini API preview Jun 24, 2025 → Vertex GA Aug 14, 2025 → subject/style tuning removed Dec 31, 2025 → all endpoints deprecated Mar 24, 2026 → Gemini API shutdown Aug 17, 2026.