GPT-Image-2.5 vs Nano Banana 2: The Default Image Engines of ChatGPT and Gemini

Most AI images people actually see come from a default, and September's two newest flagship defaults are GPT-Image-2.5 and Nano Banana 2. On September 8 OpenAI rolled GPT-Image-2.5 — the model behind ChatGPT Images 2.5, sold through the API as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst — out to every ChatGPT, ChatGPT Work, and Codex tier. Nano Banana 2 is Google's nickname for Gemini 3.1 Flash Image, the model that has been the default image generator inside the Gemini app's Fast, Thinking, and Pro modes since February and that reached general availability for developers on May 28. These are the image engines of the two most-used AI assistants on earth — OpenAI reports more than three billion images a week across ChatGPT and the images API — and both are now shipping as mature, API-accessible platforms. What separates them is no longer access, but what each is engineered to be good at.

The evidence asymmetry is the same one that runs through every GPT-Image-2.5 comparison this month: OpenAI's model is so new that no independent leaderboard has scored it as of September 9, 2026, while Nano Banana 2 has been on the Artificial Analysis boards since spring — inside the top five for both text-to-image and image editing on the September snapshot. Everything OpenAI claims for 2.5 below is vendor-reported. Google's model, by contrast, is one of the most-measured image models in production.

How each one became a default

OpenAI made ChatGPT Images 2.5 the image engine of ChatGPT by productizing it: the rollout added Sketch, where you draw a rough visual reference with @Sketch and the model works from it; Templates for common formats like posters and product shots; image comments that target edits to a specific region; and prompt sharing. Distribution did the rest — it is simply what ChatGPT uses now, across every tier including free, on desktop, mobile, and web. For most of those users the model choice is invisible.

Google reached the same place through a different route. Nano Banana 2 sits inside the Gemini app as the default generator across all three reasoning modes, and Google pushed it further out than any chat surface — into Google Search through Lens and AI Mode across more than 140 countries, into the Flow video-editing tool, and into Google Ads. It is also the rare image model that is fully multimodal on the API: developers call it through the Gemini API with response modalities for both text and image, so one model can caption, reason about, and generate images in a single turn. That embedded-in-everything distribution is the context for any quality comparison.

The capability sheets, claim for claim

Google's engineering story for Nano Banana 2 is consistency and control at scale. It holds up to five consistent characters and fourteen objects across a workflow, renders text inside images including translation and localization, grounds real-world subjects against live web and image-search knowledge, and takes aspect ratios as narrow as 1:4 and 1:8 — the shapes of banners, posters, and phone wallpapers that square-image models cannot fill. Output runs to 4K. It also exposes configurable thinking levels (Minimal and High) so a developer can trade latency against the model reasoning through a complex prompt, and a preview feature that turns video input into context-aware stills.

OpenAI's engineering story for GPT-Image-2.5 is fidelity and edit precision. The model is split into two endpoints precisely so these don't fight: GPT-Image-2.5 Flare is the fast default for everyday and high-volume generation, with latency OpenAI claims is up to 50% lower than GPT-Image-2, and GPT-Image-2.5 Sunburst accepts longer generation times for tighter edit control. OpenAI's headline claims are reference fidelity — a person, pet, or product stays recognizable when you change setting or style — plus multi-turn editing that alters only what was asked, sharper fine detail, transparent backgrounds, text rendering that handles Chinese correctly, and C2PA provenance on every output.

The two models are strongest on different axes. Nano Banana 2 is the model that keeps ten different things consistent across a whole design system; GPT-Image-2.5 is the model aimed at the edit chain where every revision has to leave everything else untouched. Those are complementary strengths, and neither vendor is chasing the other's core.

What the independent boards say about the two

Nano Banana 2 sits just inside the top five on both major Artificial Analysis image boards as of September 2026 — a consistent, well-measured performer whose strength shows up more in text-to-image and in its consistency scores than in pure editing, where it trails the current leaders MAI-Image-2.6 and GPT Image 2 (high). It is not the single best image model on any board, but it is the best-measured top-five model with the widest distribution, which is a defensible position.

GPT-Image-2.5 occupies the opposite position: every quality claim is OpenAI's, and the only third-party data point so far is a Manus speed test reporting 2–4× faster generation than GPT-Image-2. OpenAI's predecessor GPT Image 2 (high) tops the AA text-to-image board and ranks #2 in editing, so the rational prior is that 2.5 will land somewhere near or above those numbers when it appears — but a prior is not a measurement, and this is the first GPT-Image generation in a year that ships with no independent validation on day one.

The price shapes, and the one honest conversion

Google prices Nano Banana 2 in output tokens at $60 per million, which resolves to clean per-image prices: about $0.067 for a 1024×1024 image, $0.101 for 2048×2048, roughly $0.151 at the 4K end, and about half that in batch mode — with text input billed separately at $0.25 per million tokens. OpenAI prices GPT-Image-2.5 at $8 image-input / $30 image-output per million tokens, the same rate card as GPT-Image-2, but publishes no per-image figure; with the predecessor anchored near $211 per 1,000 images on Artificial Analysis's listing, an OpenAI flagship image at high quality is likely to cost several times a Nano Banana 2 image at comparable resolution. Neither price makes either model cheap at true scale, but Google's is the only one of the two you can forecast.

Because both companies run their own clouds, the routing question is real for a team that wants both. OrcaRouter carries Google's image line today — the Gemini 3.1 Flash Image preview snapshot (google/gemini-3.1-flash-image-preview) and the older gemini-2.5-flash-image — at Google's list price passed through with 0% markup, so if your stack already calls Gemini image models through one endpoint, Nano Banana 2's family is reachable without a second contract. GPT-Image-2.5 is not on OrcaRouter's catalog as of September 9, 2026 — only the previous generation GPT-Image-2 is — so the new OpenAI model means a direct OpenAI key for now.

Which default should your product default to?

• Release — GPT-Image-2.5: September 8, 2026 GA. Nano Banana 2 (Gemini 3.1 Flash Image): February 2026 launch, May 28 GA.

• Price — GPT-Image-2.5: $8/$30 per million image tokens, per-image unpublished. Nano Banana 2: $0.067 per 1K image, $0.101 per 2K, ~$0.151 per 4K; batch at half.

• Resolution and shape — GPT-Image-2.5: up to 3840×2160. Nano Banana 2: 4K plus narrow 1:4 and 1:8 ratios.

• Consistency features — GPT-Image-2.5: reference fidelity across edits (claimed). Nano Banana 2: five characters / fourteen objects, world-knowledge grounding (claimed).

• Independent scores — GPT-Image-2.5: none yet. Nano Banana 2: AA top five on both boards.

• API shape — GPT-Image-2.5: images API, two quality endpoints. Nano Banana 2: multimodal Gemini API, text and image in one call.

If your product is a design system or content engine that needs consistent characters and objects across thousands of images in unusual aspect ratios, Nano Banana 2 is the engineered answer, it is measurable, and it is the safest default in this comparison — you can reach it through Google directly or through OrcaRouter's pass-through route today. If your product is edit-heavy creative where the marginal image must be flawless and budget is secondary, GPT-Image-2.5's Sunburst endpoint is the most compelling unmeasured bet on the market. The discipline that applies to every matchup in this series applies here too: OpenAI's own predecessor GPT Image 2 (high) is still the top-scoring text-to-image model on Artificial Analysis, so GPT-Image-2.5 has to beat its own house record before it earns the right to be compared to Google's well-measured default. Watch the boards for its first appearance — that is the event that turns this comparison from claims into data.