Generated title card reading 'FIBO 2.0 vs GPT Image 2.5', subtitled 'An open 8B checkpoint against OpenAI's token-metered pair', with three chips reading '192.29 Elo gap on text-to-image', '$32 vs $210.72 per 1,000 images', and 'one is downloadable'. A footer reads 'Elo and modelled price per Artificial Analysis, AA-Image-T2I v2.1, read 20 November 2026.'
Guides & Insights

FIBO 2.0 vs GPT Image 2.5: 192 Elo, Six Times the Price

Author

Elias Hawthorne

Date Published

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

Fibo 2.0 and GPT Image 2.5 are separated by 192 Elo and a factor of 6.6 in modelled cost, and both of those numbers are softer than they look. The Elo gap is real — 1,229.27 against 1,036.98 on Artificial Analysis's text-to-image board, intervals that do not come close to touching — and it is also the least useful way to choose between them, because the two models are not trying to do the same job. Fibo 2.0 is an 8B flow-matching checkpoint that Bria released on 20 October 2026, one you can download and that expands your instruction into a structured JSON caption before drawing. GPT Image 2.5 is not one model: it is GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare, two separately snapshotted models billed per token, one of which is the strongest image generator on the public boards.

So the honest framing is not "which is better". It is "which of these four things do you need" — a downloadable model, the best output, a per-image budget, or an edit that does not drift. This page takes them in that order.

Two models, one name, and why that matters here

There is no gpt-image-2.5. Paste that string into the API and you get an error, because the identifier does not exist. What exists is Sunburst, whose ID is gpt-image-2.5-sunburst, described by OpenAI as its most capable image generation and editing model, and Flare, gpt-image-2.5-flare, positioned for fast everyday generation. Both carry the 2026-09-08 snapshot date, and there is no umbrella model that dispatches to either.

That matters in a comparison because the two variants are not close to each other. On the text-to-image board Sunburst sits at 1,229.27 and Flare at 1,209.81 — a 19-point spread between two models with the same version number. On editing the spread widens to seventeen points in the same direction, 1,201.50 against 1,184.65. Any claim of the form "GPT Image 2.5 beats X" is a claim about Sunburst, and it is worth 19 points of precision to say so.

That split matters for reading the score. Fibo 2.0 arrives in two pieces: a full Dev base that both draws and edits, and an editing-only Turbo distilled down to four steps at guidance 1.0 and explicitly not a fine-tuning base. Artificial Analysis puts a single Fibo 2.0 row on each board at the Dev checkpoints and never reports the Turbo separately, so 1,037 and 1,086 describe Dev — the slower, general checkpoint — and not the fast editor you would actually put in front of a user.

• Text-to-image — GPT-Image-2.5 Sunburst 1,229.27, rank 2 of 116. Fibo 2.0 1,036.98, rank 32. Gap 192.29.

• Image editing — Sunburst 1,201.50, rank 4 of 86. Fibo 2.0 1,085.98, rank 31. Gap 115.52.

• Modelled price — $210.72 per 1,000 images against $32.00. That is 6.6×, and it is Artificial Analysis's own comparable figure rather than either vendor's rate card.

• Weights — Fibo 2.0 downloads today, 8B parameters, CC BY-NC 4.0. GPT Image 2.5 downloads never; there is no open checkpoint in the GPT-Image line.

Screenshot of the Artificial Analysis text-to-image leaderboard, captured in English, showing GPT Image 2.6 Sunburst at the top, GPT Image 2.5 Sunburst (max) at 1,229.27 in rank 2, GPT Image 2.5 Flare (max) at 1,209.81 in rank 5, and Fibo 2.0 at 1,036.98 in rank 32, with the price column reading 210.7 for the OpenAI models and 32.0 for Fibo 2.0.

The gap that does not close

It is worth being blunt about the quality difference, because a comparison page that hedges it is useless. On text-to-image, 192 Elo is not a difference of taste. Convert it to the thing it is proxy for and it means Sunburst is preferred in something approaching nine of every ten direct comparisons. On editing the gap halves to 116 and remains outside both intervals: 1,201.50 with a range of 1,190.50 to 1,212.50, against 1,085.98 with 1,072.98 to 1,098.98. Even Sunburst's lower bound sits 91 points above Fibo's upper bound.

Nothing in the rest of this article undoes that. What the rest of the article argues is that there are four jobs where 116 or 192 Elo is not the deciding variable, and that buyers routinely pick the higher-scoring model for all four of them.

What per-token billing does to a per-image budget

Here is where the price gap gets interesting, because the two vendors do not meter the same way and the board's modelled figure hides it.

OpenAI bills image models per token — image output tokens, which is why the same model has a published input price and output price rather than a per-image rate. Through the OrcaRouter catalogue the two variants are listed at $8 per million input tokens and $30 per million output tokens, identical between Sunburst and Flare, which is the provider list price passed through with no markup on top. What that means per image depends entirely on the resolution and quality setting you ask for, and it can move by a factor of several across the settings range. The board's $210.72 per thousand is a modelling choice that assumes a representative setting, not a rate you will find printed anywhere.

Bria bills per image. $0.032 on its own platform, which makes a thousand images $32 — the same figure the board carries, which is unusual and worth noting: for Fibo 2.0 the modelled price and the unit price agree. There is a second charge if you use the VGL helper that expands a plain instruction into the structured caption, around $0.02 per call, so a full structured-prompt generation is nearer $0.052. Write the JSON yourself and you skip it.

• Predictability — Fibo's cost per image is a constant. GPT Image 2.5's is a function of your settings, which is good news if you are disciplined about resolution and bad news if a product manager keeps asking for one more pass at 4K.

• Ceiling — a thousand Sunburst images at the board's modelled rate is $210.72; a thousand Fibo images on Bria's platform is $32.00, or $52.00 with the structured-prompt call on every one.

• Floor — Fibo on your own GPU, under a non-commercial licence, is electricity. A community workflow author measured 19 seconds per image at 2K and 4 megapixels with one reference, eight steps, on an RTX 6000 Pro. There is no equivalent floor for GPT Image 2.5 at any price.

Generated cost comparison card titled 'What a thousand images costs', with two columns. The left column, 'Fibo 2.0', reads: Bria platform 32.00 dollars; with structured-prompt helper 52.00 dollars; own GPU, electricity, 19 seconds per 2K image. The right column, 'GPT Image 2.5 Sunburst', reads: modelled price 210.72 dollars per 1,000 images; input 8 dollars per million tokens; output 30 dollars per million tokens. A footer reads 'Fibo rates per Bria; modelled price per Artificial Analysis; token rates per the OrcaRouter catalogue, 20 November 2026.'

Drift, and the structured-prompt argument

The one place Fibo 2.0 makes an argument on quality rather than price is the thing its architecture is built around, and it is worth testing rather than dismissing.

Fibo 2.0 does not consume your sentence directly. A Qwen3-VL-4B-Instruct encoder reads the prompt, a Perceiver resampler compresses it into 256 to 512 gist tokens, and five of the 44 DiT blocks also receive the VLM's own layers through cross-attention. In practice the model expands "a red mug on a walnut desk, morning light from the left" into a structured VGL/JSON caption with fields for subject, material, lighting and position, generates from that document, and hands the JSON back to you. Change one field and re-run, and the fields you did not touch are still the values they were.

That is a different guarantee from "preserves detail". It means the edit is bounded by a data structure rather than by how well a language model re-reads your sentence. For multi-step product work — same object, six lighting conditions — that is the failure mode that costs the most time, and it is the one this design targets. Bria also pairs it with up to four reference images addressed in order as <image_1> through <image_4>, and a masked mode where white regions regenerate and everything else is preserved.

Whether it beats Sunburst's editing at the task level is not something the Elo board can tell you, and I would not claim it does. The board says Sunburst edits better on a general prompt distribution. The architecture says Fibo should drift less on a fixed-attribute loop. Those are different claims and only one of them has been measured publicly.

Where each one is the obvious answer

Sunburst is the obvious answer when the brief is "make it look good and nobody will ask what it cost". Marketing hero images, campaign variants, anything where a human will pick the best of twelve. Rank 2 on a board of 116 models with a 95-point lead over the best downloadable alternative is not a marginal advantage; it is the reason the model exists.

Flare is the obvious answer when you want most of that quality at lower latency and your volume is high — same price per token, nineteen Elo behind, positioned by OpenAI for everyday generation rather than showcase work.

Fibo 2.0 is the obvious answer in three narrower cases. You need the weights: for on-premises deployment, for a fine-tune, for a research baseline, for an air-gapped environment. You need a licence you can point at: Bria's pitch is licensed training data, C2PA provenance and enterprise indemnification, and it claims full EU AI Act compliance, which is a procurement answer rather than a quality one. Or you need bounded iteration on a fixed subject and the structured-prompt design is exactly the shape of your problem.

The fourth case is the one people get wrong: cost. Fibo is 6.6 times cheaper on the board, but a team that spends two weeks of engineering time working around a 192-Elo quality gap has spent more than the difference. Cheap inference on the wrong model is not cheap.

The routing answer, for the half that is ours

OrcaRouter carries GPT-Image-2.5 Sunburst and Flare — both are on the catalogue as image-generation endpoints at the provider's own $8 input and $30 output per million tokens, with no markup added. It does not carry Fibo 2.0: Bria's weights and its platform are the route to that model, and no router changes the licence attached to them. Saying otherwise would be the kind of claim that reads well and costs someone a legal review.

What one endpoint is actually worth here is the surrounding work. A Fibo pipeline has a vision-language model in it by design — the structured caption is produced by one — and that call can go through the same key as everything else, including the OpenAI models you are probably already using for the text side. Where the routing layer earns its keep is the boring failure: automatic failover means a batch of two thousand images does not stall because the captioning provider had a bad twenty minutes, and routing rules let you send cheap work to cheap models and the one hero image a day to Sunburst without maintaining two integrations and two invoices.

0% markup matters more than it sounds in a market where vendors reprice. A provider price cut shows up here the same day rather than at the next contract renewal, and the comparison you ran this month is still the comparison next month.

What would change this verdict

Two things, and only one is likely.

The first is a Fibo 2.0 fine-tune that closes the editing gap on a specific domain. Fine-tunes are where an 8B open model stops competing on general Elo and starts competing on your data, and the LoRA trainer in the repository makes that a weekend rather than a project. If a specialised Fibo beats Sunburst on your task, the general board is irrelevant and 6.6× cheaper becomes 6.6× cheaper.

The second is a change to GPT Image 2.5's cost structure. OpenAI has repriced image tokens before and may again; a cut to the output rate would compress the only axis where Fibo is comfortably ahead, and would make the "cheap enough to be worth the quality gap" argument much harder to run.

Screenshot of OpenAI's image generation documentation page, captured in English, whose opening section names the two IDs gpt-image-2.5-sunburst and gpt-image-2.5-flare and advises Sunburst where editing precision matters and Flare for fast everyday generation, followed by the Image API's Generations and Edits endpoints, the Responses API's multi-turn editing and File ID inputs, a Choosing the right API section, and a GPT-Image-2.6 paragraph on complexity control, markup-guided edits, SVG and animated GIF output.

Until one of those happens, the split is clean and unglamorous. Buy Sunburst for the images that have to be right. Download Fibo for the ones that have to be yours — yours to run, yours to fine-tune, yours to defend in a procurement meeting. Six times the price is a lot to pay for a preference, and 192 Elo is a lot to give up for a licence. Both statements are true, which is why the two models can coexist without either being the wrong choice.