Nano Banana 2.1 vs GPT-Image-2.5: One Has a Leaderboard, One Has a Label

Put Nano Banana 2.1 and GPT-Image-2.5 side by side on October 6, 2026 and the honest comparison is not between two image models — it is between a leaderboard entry and a menu label. GPT-Image-2.5 shipped on September 8, 2026 to every ChatGPT, ChatGPT Work and Codex tier, with two callable API identifiers, and Artificial Analysis added both of them to its boards three days later. Nano Banana 2.1 exists as a selectable string inside Flow's model picker and nothing else: no announcement, no model card, no identifier, no price, no score. If you came here to pick between them, the answer this week is forced, because one of them is not callable. The useful question is what would have to change for that to stop being true.

The state of each side, in one pass

GPT-Image-2.5 is a family, not a single checkpoint: GPT-Image-2.5 Flare carries the identifier gpt-image-2.5-flare and is positioned as the fast default, while GPT-Image-2.5 Sunburst carries gpt-image-2.5-sunburst and takes longer per generation for tighter edit control. Both went to the API the same day the consumer rollout started.

Nano Banana 2.1 is a 0.1 bump on Nano Banana 2, which is Google's shipping name for Gemini 3.1 Flash Image. The label was first spotted as an internal reference in the Flow web build on September 27, replacing an earlier "Nano Banana 2.5 Flash" string, and it became selectable in the interface on October 5. Google has said nothing about it since.

That asymmetry is the whole article. Everything below is either a measured number for GPT-Image-2.5 or a documented absence for Nano Banana 2.1, and the two cannot be argued against each other.

What the boards say — for the side that has boards

Artificial Analysis scored the GPT-Image-2.5 pair on September 11, 2026, three days after launch, which makes this matchup unusual for OpenAI: the 2.5 generation did not spend any meaningful time as a vendor-claims-only release.

• Text-to-image — GPT-Image-2.5 Sunburst (max) Elo 1,197 and GPT-Image-2.5 Flare (max) Elo 1,191, first and second on the board; GPT-Image-2 (high) third at 1,172.

• Image editing — GPT-Image-2.5 Sunburst (max) Elo 1,182 and GPT-Image-2.5 Flare (max) Elo 1,162, first and second; MAI-Image-2.6 third at 1,137.

• Cost as measured — Artificial Analysis lists both GPT-Image-2.5 variants at $210.70 per 1,000 images on its own price scale, against $211.00 for the GPT-Image-2 (high) they replace.

• OpenAI's token rate card — $8 per million image input tokens and $30 per million image output tokens, unchanged from GPT-Image-2, with no published per-image figure.

For Nano Banana 2.1, the corresponding rows do not exist. There is no Elo because no board has run it. There is no price because Google has published none. There is not even a resolution or an aspect-ratio list, which every other member of the Nano Banana family has documented. Any spec you read for Nano Banana 2.1 right now was inferred from its name.

Two of those rows decide the matchup on their own. The API identifier row is why a developer cannot select Google's model, and the Elo rows are why OpenAI's model no longer has to be taken on faith — the board published its numbers three days after launch, which is fast for any vendor.

The honest substitute: GPT-Image-2.5 vs the Nano Banana you can actually call

If the choice is real rather than hypothetical, the Google side of it is Nano Banana 2 — Gemini 3.1 Flash Image — because that is the model with an identifier (gemini-3.1-flash-image), a GA date (May 28, 2026, following the February 26 consumer rollout), and a scored position.

• Quality — GPT-Image-2.5 Sunburst leads the text-to-image board at Elo 1,197; Nano Banana 2 sits at Elo 1,125. On editing, Sunburst is at 1,182 against Nano Banana 2's 1,108.

• Price — GPT-Image-2.5 is OpenAI's premium image tier, measured by Artificial Analysis at roughly $211 per 1,000 images; Nano Banana 2 is listed at $67 per 1,000, with Google's own output-token rate of $60 per million resolving to about $0.045 per 512px image and $0.067 at 1024×1024.

• Surface — GPT-Image-2.5 is the default image engine of ChatGPT on every tier; Nano Banana 2 is the default inside the Gemini app, and Google has pushed it into Search via Lens and AI Mode, into Flow, and into Ads.

• Consistency features — OpenAI claims reference fidelity across sequential edits; Google claims up to five consistent characters and fourteen objects across a workflow. Both are vendor descriptions.

That is a genuine three-way spread rather than a two-way one: OpenAI's new model wins on measured quality and costs roughly triple, and Google's shipping model wins on unit economics. A Nano Banana 2.1 that closes the quality gap without moving the price line would change the shape of that trade. A Nano Banana 2.1 that merely matches Nano Banana Pro visually — which is what the earliest Flow outputs reportedly resemble — would not, because Nano Banana Pro already scores below GPT-Image-2.5 on both boards (Elo 1,102 text-to-image, 1,099 editing).

Why "we can't score it yet" is not a dodge here

Comparisons of an unannounced model usually end with a promise to update. That is not what is happening in this case, and the difference is worth stating plainly. Nano Banana 2.1 was not leaked from a research paper or spotted in a framework integration — it is visible in a shipping Google product's interface, which means Google has a build of it running in front of users. The absence of every other artifact is therefore a deliberate publication gap, not evidence that the model is far away.

The sequence so far is the same one Nano Banana 2 itself followed: consumer surface first, developer API months later, with the identifier and the GA date arriving together. On that pattern, a Flow sighting in October implies an API identifier some time after, and nothing in the sighting tells you how long the interval is.

Routing both sides through one endpoint

The practical wrinkle in this matchup is that the two vendors' image models live behind two different commercial relationships, and the gap between what each vendor has shipped makes that awkward. OpenAI's current image family is only partly reachable through aggregators — OrcaRouter carries openai/gpt-image-2 and openai/gpt-image-1.5, but not the 2.5 pair, so the new OpenAI models currently mean a direct OpenAI key. On the Google side our catalogue carries the Gemini image line, including google/gemini-3.1-flash-image-preview and google/gemini-2.5-flash-image, at each provider's list price passed through with 0% markup.

What that buys you in a comparison like this one is failover rather than a discount. When a vendor ships a point release or moves a model between tiers, a stack that calls one endpoint can route to whichever image model is currently the right one for the workload and fail over automatically if a provider degrades — without the second contract, the second key, and the migration that a single-vendor integration implies. Nano Banana 2.1 is not on OrcaRouter, because it is not on anything but a Flow picker. When it becomes a callable model, that is the path it would arrive by.

Where this leaves a decision

If your workload needs the best measured image quality today and the budget absorbs roughly $211 per thousand images, GPT-Image-2.5 Sunburst is the defensible default — it leads both Artificial Analysis boards and it is fully documented. If your workload is high-volume and price-sensitive, Nano Banana 2 at $67 per thousand is the better trade, and it is the Google model you can actually call this week.

Nano Banana 2.1 is not a third option yet. It is a scheduled possibility, and the only disciplined way to treat it is as a trigger: when Google publishes an identifier, a price and a first independent score, the three-way comparison above collapses into a real one and the trade changes. Until then, treating a Flow menu item as a competitor to a launched, scored, priced model is the kind of comparison that gets procurement decisions wrong.