The Original Nano Banana Retires: Gemini 2.5 Flash Image Shuts Down Today on the API

After thirteen months of service, one of the most consequential image models in AI history is being switched off. Google’s Gemini Developer API retires gemini-2.5-flash-image — the model the internet knows as Nano Banana — on October 2, 2026. The deprecation page is unambiguous: “deprecated and will be shut down on October 2, 2026; migrate to Gemini 3.1 Flash Image or Gemini 3.1 Flash Lite Image.”

For a model that began as an anonymous pseudonym on crowdsourced leaderboards and ended up defining conversational image editing, the retirement is quiet — a line item on a pricing page. But for the developers who built on it, today is a hard migration deadline, and the details are messier than a single shutdown date suggests.

Why Nano Banana mattered

When Google launched Gemini 2.5 Flash Image on August 26, 2025, native image editing inside a multimodal LLM was still a novelty. The model did two things exceptionally well: it preserved character consistency across edits, and it let users modify images through plain conversation — “make the jacket red,” “add a sunset,” “put this person on a beach” — without masks, layers, or inpainting tools.

The model went viral almost immediately. Before Google acknowledged it, a mysterious “nano-banana” account was posting uncanny edits on LMArena; the reveal that it was Gemini 2.5 Flash Image turned the name into a permanent brand. Google leaned into it — its successors are Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) — and the “banana” line became Google’s default answer in the image-generation price war with OpenAI and others.

The technical formula was simple and aggressive: $30 per million output tokens, with a 1K (1024×1024) image consuming 1,120 tokens — roughly $0.034 per image. That price point, combined with sub-second latency on Flash-class serving, made it the default engine for high-volume thumbnail generation, product photo editing, and the first wave of “AI avatar” consumer apps.

The date split nobody advertised

Here is the part catching teams off guard: the October 2 shutdown applies to the Gemini Developer API (AI Studio and the gemini-2.5-flash-image endpoint). Vertex AI, Google Cloud’s enterprise platform, lists the same model as available until March 15, 2027.

That split matters operationally:

• Consumer-facing apps on the Developer API lose the model today. Any request to gemini-2.5-flash-image will start failing.
• Enterprise workloads on Vertex have a five-month grace period — long enough, Google presumably calculates, for even the slowest procurement cycle to schedule a migration.
• Free-tier users actually lost access months ago: Nano Banana was quietly removed from AI Studio’s free tier earlier this year, replaced by Gemini 3.1 Flash Lite Image.

The staggered retirement follows the pattern Google established with the broader Gemini 2.5 family. The entire 2.5 series — Flash, Flash-Lite, Pro — is being discontinued “no earlier than October 16, 2026,” with exact dates set on Google’s schedule. Nano Banana’s October 2 date is simply the first shoe to drop, landing two weeks ahead of its siblings.

The three replacements, priced

Google’s migration guidance names two successors on the Developer API, with a third (Nano Banana Pro) for quality-critical work:

Gemini 3.1 Flash Image (“Nano Banana 2”) — the direct replacement. Input at $0.50 per million tokens; image output at $30 per million tokens, with resolution-scaled token counts: 1,120 tokens ($0.039) for 1K output, 1,680 tokens ($0.101) for 2K, and 2,520 tokens ($0.151) for 4K. It adds 4K output, accepts up to five reference images per generation, and holds a 32,768-token output limit. On a per-image basis at 1K, it lands within about 15% of the original’s cost while adding resolution headroom the original never had.

Gemini 3.1 Flash Lite Image — the budget tier, and the cheapest path. Standard pricing is $0.25 per million input tokens with image output at $30 per million tokens; at 1K resolution the effective cost works out to roughly $0.0336 per image — slightly cheaper than the model it replaces. Google’s own material pitches it as generating photos 2.7× faster than the Flash tier. For the high-volume workflows that made Nano Banana popular, this is the natural landing spot.

Gemini 3 Pro Image (“Nano Banana Pro”) — the premium option, at $2 per million input tokens and image output priced around $0.134 per 1K/2K image and $0.24 per 4K, with batch mode halving that. Use it when edit quality and prompt adherence matter more than unit cost.

One migration detail worth flagging: the 3.x image models charge per input image (1,120 tokens each) in addition to text tokens. Pipelines that pass multiple reference images per request will see input costs scale in ways the 2.5-era pricing didn’t.

What actually breaks today

The failure mode is blunt. Applications still calling gemini-2.5-flash-image on the Developer API receive errors, not graceful degradation. Google’s deprecation notice has been live since spring, but deprecation notices are famously unread until the day they bite.

The blast radius is hard to size precisely, but the model’s popularity guarantees it is non-trivial: fal.ai and other inference aggregators built dedicated endpoints around it at launch, thumbnail tooling and e-commerce pipelines standardized on its pricing, and a long tail of tutorial-era code still references the model string in documentation that ranks well in search. Third-party API resellers will diverge from today — some following Google’s schedule, some keeping the model alive on cached or re-routed infrastructure, with the usual reliability caveats.

The practical checklist for anyone still on the old endpoint:

1. Grep codebases for the literal string gemini-2.5-flash-image — it appears in model IDs, not just docs.
2. Decide between Flash Lite Image (cost) and Flash Image (quality) per workflow, not per codebase.
3. Re-run cost projections with per-input-image token charges included.
4. If you’re on Vertex, calendar the real deadline: March 15, 2027 — but don’t wait; the five-month overlap is a buffer, not a plan.

The bigger picture: models now have expiration dates

Nano Banana’s retirement is a small event with a large signature: it normalizes the idea that a widely-used, production-grade AI model has a thirteen-month commercial lifespan. The Gemini 2.5 family shipped in spring 2025, dominated that year’s workloads, and is now fully scheduled for removal before the leaves finish falling in 2026.

For an industry still writing multi-year enterprise procurement contracts around specific model behavior, that’s a structural mismatch. Google’s deprecation cadence — preview, GA, sunset, replacement — increasingly resembles browser release trains more than enterprise software lifecycles. The 3.x generation landing within months of the 2.5 sunset means capability continuity is preserved, but behavior continuity is not: output styling, token accounting, and pricing all shift across the boundary.

The original Nano Banana earned its place in AI history by making conversational image editing a mainstream verb. Its exit — a Tuesday shutdown notice, a five-month enterprise grace period, and a successor that costs a fraction of a cent per image — is itself a milestone: the first full lifecycle of a viral AI model, from anonymous leaderboard pseudonym to deprecated endpoint, completed in just over a year.