gemini-3.1-flash-image shuts down on October 29, 2026
Google published the deprecation in the Gemini API release notes on October 6, 2026: the model — also known as Nano Banana 2 — will be switched off on October 29, 2026, and the designated replacement is gemini-nano-banana-2.1. That is 23 days' notice for a model released on May 28, 2026. If your product calls this model ID, this page is the whole job: what changes, what it saves you, and the checklist to switch.
gemini-nano-banana-2.1What actually changes
| gemini-3.1-flash-image (old) | gemini-nano-banana-2.1 (new) | |
|---|---|---|
| Price per 1K image (standard) | $0.067 | $0.0336 |
| Price per 2K image | $0.101 | $0.0504 |
| Price per 4K image | $0.151 | $0.0756 |
| Batch API | −50% | −50% ($0.0168 at 1K) |
| 0.5K (512px) tier | $0.045 | Removed — 1K is the floor |
| Input tokens | $0.50 / 1M | $1.50 / 1M (3× — watch edit-heavy flows) |
| Text & thinking output | $3 / 1M | $7.50 / 1M |
| Reference images | Up to 14 | Up to 14 (4 characters, 10 objects) |
| Extreme aspect ratios | Tiling artifacts reported | Fixed at 2K/4K (1:4, 4:1, 1:8, 8:1) |
The headline is real — output halves at every shared resolution — but it is not a pure discount. If your workload is mostly edits fed with many reference images, the input side of your bill triples. Price both sides before you celebrate.
What the switch saves you
Output prices only. Full comparison with every image model: AI Image API Cost Calculator →
Migration checklist
- Find every call site. Search code, configs, environment variables, no-code tools and scheduled jobs for
gemini-3.1-flash-image— including preview aliases and anywhere the ID is built by string concatenation. - Swap the model ID to
gemini-nano-banana-2.1. The API surface is the same image-generation interface; no SDK migration is expected. - Map your 0.5K traffic. The 512px tier does not exist on the new model. Anything using it for cheap previews or thumbnails moves to 1K — check downstream code that assumes small files.
- Re-run your golden prompts. Output style shifts between generations (2.1 ranks higher on text-to-image and editing leaderboards, and its text rendering and wide-aspect behavior changed). Eyeball your top 20 production prompts before switching traffic, not after.
- Re-price both sides of the bill. Output halves; input tokens triple ($0.50 → $1.50/1M) and thinking output rises ($3 → $7.50/1M). Generation-only flows save ~50%; reference-heavy editing flows save less — measure your real input/output mix.
- Check thinking settings. 2.1 exposes minimal / medium / high thinking levels. Interim “thought images” are not charged; thinking tokens are.
- Re-point batch jobs. Batch stays at −50% on the new model ($0.0168 per 1K image) — for non-interactive volume it remains the cheapest configuration on either model.
- Measure latency yourself. Independent reports put 2.1's per-image time noticeably above its predecessor's. If images sit in an interactive path, benchmark before the deadline, not during it.
- Vertex AI users: verify your surface's date. Google has published different retirement dates for the API and Vertex before (the 2.5 image model differs by months between them). Check the Vertex model versions page for the date that applies to you.
- Plan for no rollback. After October 29 the old ID returns errors — there is nothing to roll back to. Fallbacks are Nano Banana Pro (quality, 4× the 1K price) or Nano Banana 2 Lite (speed, ~$0.034 at 1K). Watch error rates closely in the first week after switching.
Timeline
| May 28, 2026 | gemini-3.1-flash-image released |
| October 6, 2026 | Nano Banana 2.1 GA; deprecation published in Gemini API release notes |
| October 29, 2026 | gemini-3.1-flash-image shuts down |
One more date for the diary: the older gemini-2.5-flash-image is listed with a March 15, 2027 shutdown on Vertex AI — if you still call that ID anywhere, its clock is also ticking.