GPT-Image-2.6 vs Nano Banana 2.1: The Leaderboard Disagrees With Itself

On 5 November 2026 the two most interesting image models you can call are a month apart in age, within about a cent of each other per picture, and neither of them is the model the independent boards rank first. GPT-Image-2.6 shipped from OpenAI on 4 November. Nano Banana 2.1 shipped from Google on 6 October. Artificial Analysis, which has scored both companies' previous generations into the ground, has no row for GPT-Image-2.6 at all — while Nano Banana 2.1 sits fourth on its text-to-image board at 1,175 Elo from 6,322 votes and fourth on its editing board at 1,151 from 7,009 votes, behind three models that are all older than the one OpenAI announced this week.

That is the whole problem with this matchup in a paragraph. The model with a month of votes is the cheaper, slower-moving, already-shipped one. The model with no votes is the one with new output formats. Anyone telling you which of these is better is either quoting a vendor or guessing.

Two launches, two opposite instincts

Google's October release is a price-and-plumbing move. Nano Banana 2.1 — API ID gemini-nano-banana-2.1, built on Gemini 3.6 Flash — did not add a new output format, a new medium, or a new kind of control. It cut image output tokens from $60 to $30 per million, which halves the cost of a 1K render to $0.0336 and a 2K render to $0.0504, and left 4K at $0.113 because the higher resolution carries 3,780 image tokens rather than 2,520. It added a named resolution ladder, a fourth reference slot count if you were counting, and search grounding. The pitch is predictability: you know the tier, you know the token count, you know the bill.

OpenAI's November release is the other instinct entirely. GPT-Image-2.6 held the rate card exactly where GPT-Image-2.5 left it — $5 per million text input tokens, $8 per million image input tokens, $30 per million image output tokens — and spent the release on new dials and new deliverables: a complexity control, markup-guided editing, SVG export, animated GIF output, transparent backgrounds, custom sizes in multiples of 16 pixels. The pitch is capability: here are things the model can make that the previous one could not.

Read those two paragraphs again and you can predict almost every practical difference between the products. One company is optimising the unit economics of a thing that already works. The other is expanding the set of things the thing can be.

What the independent boards actually say

Artificial Analysis runs two separate image boards and the separation matters — a model can be excellent at rendering a prompt from nothing and mediocre at following an edit instruction, and averaging the two hides exactly the difference a production team cares about.

• Text-to-image — GPT-Image-2.5 Sunburst (max) 1,224 Elo, 13,153 votes, first. GPT-Image-2.5 Flare (max) 1,207, second. GPT Image 2 (high) 1,188, third. Nano Banana 2.1 1,175, 6,322 votes, fourth. GPT-Image-2.6: no row.

• Editing — GPT-Image-2.5 Sunburst (max) 1,200, 18,370 votes, first. MAI-Image-2.7-Preview 1,197, second. GPT-Image-2.5 Flare (max) 1,183, third. Nano Banana 2.1 1,151, 7,009 votes, fourth. GPT-Image-2.6: no row.

• Against its own predecessor — Nano Banana 2.1 leads Nano Banana 2, which sits seventh on text-to-image at 1,140 Elo from 16,224 votes, by 35 points.

Two caveats before anyone builds a purchasing decision on those numbers. The first is sample size: Nano Banana 2.1's 6,322 text-to-image votes are less than half of GPT-Image-2.5 Sunburst's 13,153, so its confidence interval is wider and its position is less settled than the ranking implies. The second is that a 49-point gap between first and fourth on a 1,000-point-ish Elo scale is real but not enormous — it is the kind of difference that shows up reliably in blind pairwise votes and intermittently in whether a specific image is usable for a specific job.

Two vendor claims that cannot be compared to each other

Both companies publish their own human-preference numbers and both sets are unusable for this decision, for different reasons.

Google reports Nano Banana 2.1 at 1,050 Elo in thinking mode and 1,015 without thinking on an internal human side-by-side eval, against 990 for Nano Banana 2 and 935 for Nano Banana Pro. Those are real measurements of real raters, and they are on Google's own scale with Google's own prompt distribution — which is why 1,050 and Artificial Analysis's 1,175 both describe the same model without contradicting each other. Do not compare across them.

OpenAI cites Design Arena for a first-place placing for GPT-Image-2.6 Sunburst. Same problem in a different shape: a different board, a different rater pool, and a vendor selecting the board that flatters the release. It may well be true. It is not a number you can put next to 1,175 and subtract from.

The only honest summary is that Google has an independently measured position and OpenAI does not have one yet. That asymmetry should shape how much confidence you extend to each.

The meters: a named ladder against a token budget

This is where the two products diverge in a way that matters more to a finance team than to a designer, and it is the difference most comparisons miss.

• 1K render — Nano Banana 2.1 $0.0336, fixed at 1,120 image tokens. GPT-Image-2.6 at medium quality roughly $0.043, estimated from about 1,449 output tokens.

• 2K render — Nano Banana 2.1 $0.0504. GPT-Image-2.6 has no named 2K tier; you specify pixel dimensions in multiples of 16 within a 1:3 to 3:1 aspect range.

• 4K render — Nano Banana 2.1 $0.113, from 3,780 image tokens. GPT-Image-2.6 publishes no 4K tier.

• Quality ceiling — GPT-Image-2.6 at max quality runs roughly $0.21 per image; Nano Banana 2.1 tops out at the 4K price.

• Input — Nano Banana 2.1 $1.50 per million tokens; GPT-Image-2.6 $5 per million text and $8 per million image.

• Batch — Nano Banana 2.1 halves at 50% of standard. GPT-Image-2.6 is not yet available through the Batch API.

• Levers — Nano Banana 2.1 changes cost by resolution tier and thinking level. GPT-Image-2.6 changes cost by quality, by complexity, and by how much content the picture contains.

Per-image estimates for GPT-Image-2.6 deserve a flag: OpenAI does not publish a fixed price per image, because it bills tokens and the token count varies with what you asked for. The figures above are derived from output-token counts, not quoted from a price list, and different outlets have published materially different numbers by assuming different token budgets. Treat any "GPT-Image-2.6 costs X per image" claim, including ours, as an estimate with a range attached.

The practical consequence is straightforward. With Nano Banana 2.1 you can forecast a month of 2K product shots to within a rounding error, because resolution determines token count determines price. With GPT-Image-2.6 you can forecast the rate but not the budget, because the complexity parameter defaults to high and planning tokens are billed at the same $30 per million as rendered ones. If your finance team needs a number that does not move, that is a real argument for Google. If your team would rather have the capability and measure afterwards, it is not.

Editing: fourteen reference slots against a saved library

Nano Banana 2.1 takes up to 14 reference images in one request — four characters and ten objects — and supports mask-based local edits across 1K, 2K and 4K. Google's own numbers put multi-character consistency at 1,106 Elo in thinking mode against 978 for Nano Banana 2, which is the largest single-generation jump in its published table and the one that maps most directly to a real workflow: keeping the same three people consistent across a campaign.

GPT-Image-2.6's editing story is about sequence rather than breadth. OpenAI describes Sunburst as more consistent across successive edits — the tenth edit drifting less than it used to — and adds markup-guided editing, where you draw on a transparent PNG overlay rather than supplying a mask. The two cannot be combined: it is markup or mask, not both. Separately, the reusable reference library — save a person or character once and reuse it — is a ChatGPT feature, not an API one, so a team building on /v1/images/edits does not get it.

If your edit workflow is "hold fourteen assets consistent in one generation", Nano Banana 2.1 is the only one of the two that does it. If it is "make twenty small changes to one image without it degrading", 2.6 is aimed squarely at you, and there is no independent measurement of how well it works.

Formats: the one place with no contest at all

GPT-Image-2.6 exports SVG and outputs animated GIF. Nano Banana 2.1 outputs raster images and nothing else. If the deliverable is a vector asset for a design system or a short looping animation for a product surface, the comparison is over before it starts — and it is worth saying plainly that this is a capability gap, not a benchmark gap, so it does not need independent verification to be true.

The reverse holds on resolution. Nano Banana 2.1 sells named 1K, 2K and 4K tiers and supports aspect ratios up to 8:1 at the top two, which means you can write "4K, 8:1 banner" into a spec sheet and get it. GPT-Image-2.6 takes custom dimensions in multiples of 16 within 1:3 to 3:1 and publishes no named resolution ladder, so a print vendor asking for a contract-specified 4K deliverable is asking you a question OpenAI's documentation does not answer.

Where Google's own model card admits the weakness

Credit where it is due: Google's model card for Nano Banana 2.1 is unusually forthcoming. It lists blurry small text at 1K, weak handling of long text passages, imperfect character consistency, occasional left-right spatial confusion, and limited 3D reasoning and factual accuracy. Those are the failure modes that hurt a real project — a product label with unreadable fine print, a diagram with the arrow pointing the wrong way.

The temptation is to say GPT-Image-2.6 does not have those problems. It might not. But there is no independent evaluation of 2.6's text rendering or spatial reasoning, and the honest position is that Google has documented its weaknesses and OpenAI has not documented any. Absence of a published weakness is not evidence of absence, particularly for a model released 24 hours before this was written.

Picking one, given all of that

• If you need print-resolution or large-format output with a budget you can sign off in advance, take Nano Banana 2.1. The named ladder, the 8:1 aspect support and the batch discount are all things 2.6 does not currently offer.

• If your deliverables include vector or motion, take GPT-Image-2.6. No amount of price advantage on the other side buys you an SVG.

• If your work is long edit sequences on a small number of assets, 2.6 is the better-targeted product, with the caveat that nobody outside OpenAI has measured whether it delivers.

• If what you want is the best independently verified still image today, take neither. That is GPT-Image-2.5 Sunburst, at 1,224 on text-to-image and 1,200 on editing, and it is a month older than one of these models and two months older than the other.

That last bullet is the one people will skip past, and it is the most useful sentence here. Both vendors are selling you a step forward. On the only public evidence that was not produced by either of them, the step forward is not yet visible.

How to see both without betting on either

OrcaRouter does not carry GPT-Image-2.6 or Nano Banana 2.1 as of 5 November 2026, so this is not a piece where we can tell you to point one endpoint at both. What we do carry is the previous generation of each — GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare, and Nano Banana 2 as gemini-3.1-flash-image-preview — on one key, at list price with 0% markup so the provider's rate is the rate you pay, with automatic failover between providers.

Which happens to be the right way to start anyway. The models with verified scores are exactly the models you can call today from a single endpoint and a single bill, and running your own prompts through those two tells you which vendor's aesthetic and failure modes fit your work — before you commit to either vendor's newest release on the strength of a launch post. When each new model reaches the catalogue, the pass-through rule means the vendor's price arrives with it unchanged, so the only thing that changes on your side is the model string.

That is the sequence the evidence supports: settle the vendor question on the generation that has been voted on, then move to the new one when there is a number to move on. The alternative — picking between two models whose only comparable scores belong to their predecessors — is a bet dressed up as a comparison.