A generated title card headlined 'FIBO 2.0 vs Nano Banana 2.1', subtitled 'Thirty-two dollars and thirty-three sixty; sixty-six Elo'. Two price tags read '$32.00 per 1,000 images' and '$33.60 per 1,000 images', with a bracket between them reading '66 Elo on editing'. A footer reads 'Prices and Elo per Artificial Analysis, AA-Image-Editing, read 20 November 2026.'
Guides & Insights

FIBO 2.0 vs Nano Banana 2.1: Same Price, 66 Elo Apart

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Two image models, five percent apart on price, sixty-six Elo apart on the only blind measure either of them has. FIBO 2.0 costs $32.00 per thousand images. Nano Banana 2.1 costs $33.60. On Artificial Analysis's editing board, read on 20 November 2026, Nano Banana 2.1 sits at 1,152 Elo in tenth place with 7,606 battles and an interval of 1,140 to 1,164; FIBO 2.0 sits at 1,086 in thirty-first with 5,753 battles and an interval of 1,073 to 1,099. On text-to-image the gap widens to 139 points: 1,176 against 1,037.

So the pricing question that dominates most of these comparisons is dead on arrival here. You are not choosing between cheap and expensive. You are choosing between a closed model from Goog​le's image line that is measurably better at the thing both are sold to do, and an 8B open checkpoint from Bria that costs a rounding error less per image and comes with a non-commercial licence. Everything else is a detail about which of those two facts you can work with.

What sixty-six Elo is, in practice

Elo gaps get quoted as though the number itself were the argument. It is not. Sixty-six points with non-overlapping intervals means the arena, given a large sample of side-by-side comparisons, preferred Nano Banana 2.1 often enough that sampling noise does not explain the ordering. What it does not tell you is which kinds of edits it preferred, and that is the thing you would need to know before betting a workflow on it.

A screen capture of the Artificial Analysis AA-Image-Text-to-Image leaderboard ranked by Elo, showing Nano Banana 2.1 at rank 8 with 1,176 Elo, an interval of 1,164 to 1,188 across 6,793 battles at $33.60 per 1,000 images, and Fibo 2.0 at rank 32 with 1,037 Elo, an interval of 1,026 to 1,048 across 5,741 battles at $32.00 per 1,000 images.

The pattern in the surrounding rows is suggestive. Nano Banana 2.1 is a generalist — it appears on both boards at a high rank, eighth on text-to-image and tenth on editing, with a 16,000-plus battle history across a family that includes Nano Banana 2 and Nano Banana Pro. FIBO 2.0 is at its best relatively on editing (1,086, thirty-first) and at its weakest on text-to-image (1,037, thirty-second), which is the profile of a model trained hard on instruction-following within an existing frame rather than on conjuring frames from nothing. If your work is "change this, keep the rest", the two are closer than sixty-six points suggests. If your work is "make me a picture", the gap is real and it is not close.

• Editing Elo — Nano Banana 2.1 1,152 (10th, 7,606 battles); FIBO 2.0 1,086 (31st, 5,753 battles).

• Text-to-image Elo — Nano Banana 2.1 1,176 (8th, 6,793 battles); FIBO 2.0 1,037 (32nd, 5,741 battles).

• Price per 1,000 images — $33.60 against $32.00, a 5% difference.

• Weights — none, closed API; 8.1B parameters, gated download, CC BY-NC 4.0.

• Generation and editing — one hosted model covering both; one checkpoint covering both, the first time Bria has unified them.

• Prompt format — natural language plus image references; VGL, a structured JSON caption schema, with a companion package that writes it.

A generated scoreboard card titled 'FIBO 2.0 vs Nano Banana 2.1 — the scoreboard'. Left column 'FIBO 2.0': 'Editing Elo: 1,086 (31st)', 'Text-to-image: 1,037 (32nd)', 'Weights: 8.1B, gated, CC BY-NC 4.0', 'Price: $32.00 / 1k images', 'References: up to 4', 'Prompt: VGL JSON schema'. Right column 'Nano Banana 2.1': 'Editing Elo: 1,152 (10th)', 'Text-to-image: 1,176 (8th)', 'Weights: none, closed API', 'Price: $33.60 / 1k images', 'References: multi-image, per Google', 'Prompt: natural language'. A footer reads 'Elo per Artificial Analysis, read 20 November 2026; prices per each vendor; licence per Bria model card.'

The five percent that is not five percent

Bria's $0.032 per image is flat. So is the $33.60 on the leaderboard for Nano Banana 2.1, at least as the board reports it, but a Goog​le image endpoint's real invoice depends on output size and on how many reference images you attach — the sort of thing that never appears in a per-thousand listing. That cuts both ways in this comparison and it is worth being explicit that neither number should be treated as a quote.

What is more useful is to notice how small the price difference is relative to what else changes. At a million images a year the gap is roughly $1,600 — less than a week of one engineer's time, and far less than the cost of rebuilding a pipeline around a different model's prompt format. The price column is not where this decision lives. It is there to stop you from assuming it does.

The licence is the actual price

FIBO 2.0's $0.032 buys you a commercial API call and, separately, a gated download under CC BY-NC 4.0. The download is non-commercial: Bria's card is explicit that commercial use of the weights requires a separate agreement, and the Hugging Face gate asks for a company name and an intended-use category before it will hand over the files. The API, by contrast, has no such restriction — you can sell the output of the endpoint. The two are different products that share a name.

Nano Banana 2.1 has no weights to argue about, which removes the licence question entirely and replaces it with a dependency question. You cannot audit it, cannot fine-tune it, cannot pin a build and run it in your own network, and cannot keep customer photographs from leaving your infrastructure. For a lot of companies that last clause is the whole decision, and it is decided against the model that scores ten places higher.

This is the point at which the sixty-six Elo stops mattering and the sentence "we need it inside our VPC" starts. It is also the point at which the open-weights pitch usually overpromises: a non-commercial licence is not a way to run a product in your own network, it is a way to research and prototype in your own network while you negotiate. Anyone reading "open weights" as "we can ship this" is about to have an unpleasant conversation with a lawyer.

A third option the price column hides

Both vendors sell endpoints at roughly three cents an image. That is cheap enough that the interesting question is no longer which model to buy but how to keep the surrounding pipeline from costing more than the render.

The pipeline around an image model is mostly text. A caption gets written, a schema gets validated, a render comes back wrong, a retry gets composed, a batch gets routed to a different model because one vendor is throttling. Each of those steps is a language-model call with its own vendor and its own key. OrcaRouter puts more than 200 text models behind a single key at provider list price with 0% markup, so the captioner can be swapped per job without a new contract and a provider outage gets covered by automatic failover rather than by a pager. Neither FIBO 2.0 nor Nano Banana 2.1 is on our catalogue — the image calls stay where they are — but the half of the workflow that decides whether the image call is worth making does not have to be a second integration.

For a FIBO pipeline specifically, that framing is not decorative. The model is documented against VGL, a structured caption format that separates subject, composition, lighting, camera, style and text-in-image into named fields rather than one run-on sentence. The natural author of a VGL object is a model, not a person, and the natural failure mode is a schema violation two seconds before the render. Having the writer, the validator and the retry behind one endpoint with a routed fallback is the difference between a pipeline and a pile of scripts.

Which one, and what would flip it

If you have no constraint beyond "make good images cheaply", Nano Banana 2.1 wins on the evidence, and it wins by more than the price difference — twenty-one places on editing and twenty-four on text-to-image for five percent more money is not a close call.

A generated card titled 'What five percent buys', showing a see-saw: on one side '$1,600 a year at a million images' and on the other 'a pipeline rebuilt around a different prompt format, a fine-tune you cannot run, and a licence you cannot read'. A footer reads 'Prices per Artificial Analysis, read 20 November 2026; the annual figure is arithmetic on a $0.0016 per-image difference.'

If you have a constraint — data residency, a fine-tune on your own catalogue, an audit requirement, a research programme — FIBO 2.0 is the one that can satisfy it, and the sixty-six Elo is the tax you pay for the file. The fine-tuning story is real: the repository ships a single-GPU LoRA trainer for editing and text-to-image, and because generation and editing share one set of weights, one fine-tune can teach a house style and the edits within it. That is a capability the hosted model does not offer at any price.

Three things would change the verdict. A commercial licence from Bria would make the download usable and turn FIBO into a genuine in-house option rather than a research toy. A published per-image price for Nano Banana 2.1 that accounted for references and 4K output would tell you whether the five percent is really five percent — the direction of that answer is not obvious. And a run of arena battles that separated the two editing intervals would settle whether sixty-six points is the gap or whether the bands, which currently clear each other by 41 points, are closer than they look.

Until then, the accurate summary is short. Two models, one price, one of them better on every board either appears on, and the other one the only member of the pair you can hold in your hand. That is not a leaderboard question and it never was.