
FIBO 2.0 vs HY-Image 3.5: One Board Calls It a Tie
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 514 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 116 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Put Fibo 2.0 and HY-Image 3.5 on the editing board and you get a tie. Put them on the text-to-image board and you get a 65-point gap. Same two models, same week, same methodology, opposite conclusions — and the reason is not that either score is wrong. It is that the two models are not equally good at the two things an image model does, and a single "which is better" answer destroys exactly the information you need. Tencent's HY-Image 3.5, which the vendor ships as a preview under the identifier hy-image-v3.5-preview inside its HunyuanImage line, edits roughly as well as Bria's Fibo 2.0 — the open 8B checkpoint Bria released on 20 October 2026 — and generates substantially worse. If your work is editing, that sentence is the whole article.
A note on provenance before the numbers do any work. The two Elo figures and their intervals come from Artificial Analysis, which runs blind pairwise preference voting and publishes the spread around each score; I read both boards on 20 November 2026. The two prices are that board's modelled cost per thousand images — a comparable synthetic figure, not either vendor's rate card. Tencent's own documentation and Bria's model card are the source for everything about limits, licences and release state, and where a vendor states something I have not independently checked, the sentence says so.
Reading two boards that disagree
The disagreement is stark enough to state as numbers first and interpret afterwards.
• Editing — HY-Image 3.5 1,091.87 (interval 1,078.87 to 1,104.87), rank 28 of 86. Fibo 2.0 1,085.98 (1,072.98 to 1,098.98), rank 31. HY-Image leads by 5.89 points, which is less than half of either interval half-width.
• Text-to-image — Fibo 2.0 1,036.98 (1,025.98 to 1,047.98), rank 32 of 116. HY-Image 3.5 972.12 (960.12 to 984.12), rank 75. Fibo leads by 64.86, and the intervals are 41 points apart at their closest.
• Price — $32.00 per 1,000 images for Fibo 2.0, $24.00 for HY-Image 3.5. A difference of eight dollars, or 25%.
• Release state — Fibo 2.0 is generally available with downloadable weights under CC BY-NC 4.0. HY-Image 3.5 is still labelled preview, with no second stable identifier published beside the preview one.
The editing tie is the interesting one, because it is not a tie between two models that arrived at the same place. It is a tie between two quite different bets. HY-Image 3.5 is a Chinese lab's image model served through Tencent Cloud with strong Chinese-language comprehension, built for a market where the prompt is as likely to be 中文 as English. Fibo 2.0 is a licensed-data checkpoint whose entire design premise is that prompts should be structured documents rather than sentences. They converge on the editing board and diverge by 65 points on generation, and no one who looked only at the editing board would have predicted that.

Why the generation gap is the bigger fact
Sixty-five Elo is the sort of difference that shows up in the first twenty outputs, not the two-hundredth. It is roughly the gap between rank 32 and rank 75 on the same board, which is to say it spans most of the middle of the field — the distance between a model that holds its own against Qwen-Image-2.1 and FLUX 3 Image Klein and one that sits below Wan 2.7 Pro and HiDream-O1-Image.
The editing board tells a different story because editing is a different task. An editor starts from an image that already carries most of the composition, lighting and structure. The model's job is a local change, and a model with mediocre generation can still be a competent editor if it preserves what it is given. That is exactly what the numbers describe: HY-Image 3.5 is a decent editor built on a weaker generator, and Fibo 2.0 is a strong generator whose editing is roughly as good.
The practical consequence is that the two models are not substitutes. If you are choosing one model for a pipeline that both generates and edits, the generation board should decide it, and Fibo 2.0 wins that by a margin no interval argument touches. If you have a library of existing images and you only ever edit them, HY-Image 3.5 is the cheaper of the two and statistically level on the axis you care about.
The preview label is not a formality
Tencent ships this line preview first and official version later, and the current generation is in the first of those two states. The identifier in circulation is hy-image-v3.5-preview. There is no second, stable, non-preview identifier sitting beside it, and Tencent has published no date for when one will appear.
For a comparison this matters more than it does for a review, because it changes what "cheaper" means. A preview model can be repriced, renamed or superseded without notice, and a pipeline pinned to a preview identifier is pinned to something the vendor has explicitly declined to promise. Fibo 2.0 has the opposite problem in a different register: stable weights, unstable licence. Its checkpoints will not change under you; what you are allowed to do with them is the part that requires a contract.
That is the choice in one line. A preview API is stable in licence and unstable in existence. An open checkpoint is stable in existence and restricted in licence. Neither is the safe option, and which risk you would rather carry depends on whether your problem is a legal review or a production incident.
What each model actually does
Fibo 2.0 is a flow-matching model of eight billion parameters that denoises inside the FLUX.2 VAE's latent space, so a prompt reaches the drawing stage through a Qwen3-VL-4B-Instruct encoder, a two-layer Perceiver resampler that squeezes it into 256 to 512 gist tokens, and a 44-block single-stream DiT — five of whose blocks take VLM layers directly through cross-attention rather than reading a caption. That wiring is the reason the structured-caption premise works at all. Dev is the 50-step, guidance-5.0 checkpoint and handles both generation and editing. Edit Turbo is the distilled sibling: four steps, guidance 1.0, editing only, and not a base anyone should try to fine-tune.
• Reference images — Fibo 2.0 accepts up to four, addressed in order as <image_1> through <image_4>. HY-Image 3.5 accepts up to five, per Tencent's documentation.
• Output envelope — Fibo 2.0 produces multiples of 16 at roughly one megapixel, defaulting to 1024×1024, 1424×800 or 1248×832. HY-Image 3.5 is documented to 2K.
• Structured prompts — Fibo's differentiator. A short instruction becomes a JSON/VGL caption with fields for subject, material, lighting and position, and you get the JSON back. Change one field, re-run, everything else holds.
• Chinese comprehension — HY-Image 3.5's differentiator, and the reason it exists as a separate product rather than a localisation of one. Tencent's own product material pitches strong Chinese comprehension alongside image generation.
• Access — Fibo 2.0 through Bria's platform, its licensing programme, ComfyUI and the community ecosystem, and Civitai's generator since 18 November. HY-Image 3.5 through Tencent Cloud's TokenHub API surface.

Cost, once you stop trusting the per-image number
Eight dollars per thousand is the headline and it is the wrong headline. Two things change the arithmetic.
The price on Bria's platform is $0.032 an image, and here the board's modelled $32 happens to land on it exactly. What neither number includes is the structured-prompt helper: the separate vision-language call that rewrites your instruction as a JSON caption, billed at roughly $0.02 each time. Run it on every render and the true cost drifts to $0.052, or $52 a thousand — better than double HY-Image 3.5's $24. Skip the helper, hand-write the JSON, or reuse one cached caption across a fixed scene and change only the fields that move, and $32 is what you actually pay.
The second is that HY-Image 3.5's $24 buys you a preview. If the model is repriced on general availability — and preview pricing is usually introductory — the comparison you ran in November may not survive to February. Fibo's $32 is a published rate on a shipped product, and the weights are free for non-commercial use if you would rather pay in GPU hours.
• Cheapest per image — HY-Image 3.5 at $24 per thousand, with preview pricing risk.
• Cheapest with the helper on every call — HY-Image 3.5 by more than 2×.
• Cheapest at scale, if you own hardware — Fibo 2.0, at zero marginal cost under a non-commercial licence. One community workflow measures about 19 seconds per 2K image on an RTX 6000 Pro.
• Cheapest for a commercial product on a hosted API — neither, straightforwardly. Fibo needs Bria's agreement; HY-Image's preview terms are the thing to read, and they are on Tencent Cloud rather than in a marketing page.
Where OrcaRouter fits — and where it does not
Say the awkward part first: neither model is on the OrcaRouter catalogue. Fibo 2.0 is a Bria checkpoint under a Bria licence; HY-Image 3.5 is a Tencent Cloud product. Neither is something a router can hand you, and pretending otherwise would be the one useful thing this page could do wrong.
The part of this pipeline that a router can hold is everything except the pixels. A Fibo render starts with a caption written by a vision-language model and can end with a second one grading whether the edit left the rest of the frame alone; both are text calls, and both can sit on a single key beside 200-plus other models. That matters most when a batch is long: routing rules put the cheap prompt-rewriting on a cheap model and keep the careful grading on an expensive one, 0% markup means whatever the provider charges that day is what you pay, and automatic failover keeps a two-thousand-image run alive when one provider's endpoint goes soft for an hour.
If you specifically want HY-Image 3.5's Chinese comprehension, note that Tencent's own language models are a different story and a different product line — check the catalogue for what is actually there rather than assuming an image model implies its sibling is routable.
Who should pick which
Start with the task, not the score, and the answer usually falls out.
If you generate as much as you edit, choose Fibo 2.0. The 65-point generation gap is the largest number in this comparison and it is the one that will show up in your output every day. HY-Image's 5.89-point editing lead is inside the noise and will not.
If you only edit existing images, and your prompts are in Chinese, choose HY-Image 3.5. It is level on the board, cheaper per image, and Chinese comprehension is the one capability Fibo 2.0's documentation does not claim.
If you need to own the model, there is no choice to make. HY-Image 3.5 is closed and preview. Fibo 2.0 downloads, fine-tunes, quantises and runs air-gapped — for non-commercial use, or commercially under an agreement.
If you are procuring for a regulated buyer, the tie on the editing board is irrelevant and the question is which vendor will sign what. Bria sells indemnification and provenance as the product. Tencent sells a cloud service. Those are different procurement conversations and only one of them ends with a document your legal team likes.

The last thing worth saying is about how to read a comparison like this at all. Two models, four numbers, and the conclusion flips depending on which pair you look at. That is not a failure of the benchmarks — it is the benchmarks doing their job, which is to tell you that "better" is not a single axis. Pick the axis your work actually lives on, check whether the gap on that axis clears its own error bars, and ignore the rest of the table.
