Hy Image3.5 vs Meta Muse Image: One Draws, One Goes and Finds Out

Meta Muse Image and Tencent Hunyuan Hy Image3.5 preview landed seven weeks apart, and they are not really competing for the same job. Muse Image, released 7 July 2026 out of Meta Superintelligence Labs under the internal codename Mango, is an agentic image model: it searches the web, runs code to build the thing it is drawing — QR codes, charts, infographics — and corrects itself mid-generation. Hy Image3.5 preview, live since 22 September 2026, is a straightforwardly excellent-looking generator and multi-turn editor priced at ¥0.15 per 2K image. If your prompt is "make me an image", both are fine. If your prompt is "make me an image of today's actual weather in Lisbon, as a chart", only one of them is even playing.

That difference is why this comparison is more interesting than the price columns suggest, and also why the price columns are less decisive than they look.

What each one is built to do

Hy Image3.5 is Tencent Hunyuan's native multimodal image model, served as hy-image-v3.5-preview. Text-to-image and image-to-image in one model, multi-turn editing that keeps earlier turns as context, up to five reference images per request, standard aspect-ratio coverage, 2K output ceiling. It is charged per delivered output image, ¥0.15 at 2K on Tencent Cloud in China and $0.024 per image on Tencent Cloud International, with failed and discarded generations unbilled. It reached preview after roughly two months of gray-scale testing inside Yuanbao, during which Tencent says image-generation request volume there rose more than 50%.

Meta Muse Image is the opposite kind of product. It shipped first as a consumer feature rather than an API: free in the Meta AI app and on meta.ai, powering Instagram Stories AI effects in the US, live in WhatsApp direct chats in select countries, with Facebook and Messenger planned and a route into Meta's Advantage+ Creative advertising tooling. It carries an invisible watermark called Content Seal that Meta says survives cropping, resizing and recompression. It also lets users tag public Instagram accounts in generated images without notifying them, with an opt-out setting — a design decision worth knowing about before you put it anywhere near a brand.

• Price — Hy Image3.5 $0.024 per 2K image, billed on delivered output only; Muse Image is free in Meta's own apps and listed by Artificial Analysis at $10 per 1,000 images as an API-equivalent

• Capability class — Hy Image3.5 generates and edits from reference images; Muse Image generates, searches the web, executes code and self-corrects

• Multi-turn editing — Hy Image3.5 keeps prior turns as context across a session; Muse Image's conversational editing is inside the Meta AI assistant

• Reference images — Hy Image3.5 up to five per request; Muse Image works alongside the Muse Spark text models for layout and reasoning

• Output ceiling — Hy Image3.5 up to 2K; Muse Image is tuned for social formats rather than print-resolution output

• Independent placement — Muse Image Elo 1,112 at #7 of the Artificial Analysis text-to-image board, 7,344 samples, and 1,116 at #6 on the editing board; Hy Image3.5 has no public arena entry

• Access — Hy Image3.5 via Tencent's apps and cloud; Muse Image inside Meta's consumer surfaces, not as a general API

The free one is not the cheap one

Muse Image being free inside Meta's apps is a distribution decision, not a pricing one, and it does not survive contact with a production pipeline. There is no general-purpose public API to call, no per-image rate card to budget against, and no SLA attached to a feature inside a consumer app. The $10-per-1,000-images figure on Artificial Analysis is a derived API-equivalent, useful for comparing capability per dollar on the board and not a quote you can go and buy at.

Hy Image3.5, by contrast, is a priced, documented cloud service from the day it previewed — Tencent Cloud TokenHub, MPS and VOD on the platform side, Yuanbao, ima, Miora, WorkRally, OnSolo and WorkBuddy on the application side, with a free trial on three of those running from 22 September to 7 October. If you are shipping something, that is the difference between a product you can build on and a product you can screenshot.

What the independent board says about the Meta side

Artificial Analysis has Muse Image at Elo 1,112, seventh on the text-to-image board from 7,344 samples collected in July 2026 — the cheapest entry of any model in the top ten at a derived $10 per 1,000 images. On the editing board it does better still at 1,116, sixth overall, ahead of Nano Banana 2, ahead of MAI Image 2.5 Pro and ahead of both Qwen-Image-3.0 variants. For a model that was a consumer launch rather than an API launch, that is a strong independent result.

Tencent's evidence for Hy Image3.5 is a blind GSB preference study run with hundreds of professional designers inside Tencent, reporting about a 30% gain over Hy Image3.0, parity with Seedream 5.0 Pro, and a marginal edge over Nano-Banana Pro and Qwen-Image-3.0 Pro. Vendor-run, vendor-populated, no published prompt set, and as of writing the only performance evidence that exists for the model. There is no Artificial Analysis entry, no arena placement, and no third-party reproduction of any of it.

The Meta models we do route are not this one

Worth being precise here, because it is exactly the kind of thing a model page can blur. OrcaRouter serves Meta's Muse Spark text models, meta/muse-spark-1.1 and meta/muse-spark-1.2. We do not serve Muse Image. Hy Image3.5 is not on OrcaRouter either — we host no Tencent image model, and the only Tencent entry in our catalogue is the text-only Hy3 line — so reaching it means Tencent's own cloud and several third-party platforms.

Where OrcaRouter earns its place in a pipeline like this is on the production side. Image models we do serve — openai/gpt-image-2, google/gemini-3.1-flash-image-preview, grok/grok-imagine-image, the Imagen 4 family — sit behind one OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor price change is live the same day, with automatic failover between routes and a routing DSL that lets you pin a specific model per request instead of wiring an SDK for each vendor. When you are deciding whether to move your image spend onto a model that previewed this week, having the incumbent lane already wired is what makes the test cheap.

What to actually do

If your image work is consumer-facing, social-formatted and benefits from the model looking things up or drawing a chart from real data, Muse Image is doing something Hy Image3.5 does not attempt, and it is free at the point of use inside Meta's apps — but you are building inside someone else's product surface, with no API and no commitment. If you need a priced, billable, multi-turn image service with reference-image conditioning and 2K output at a rate that makes high-volume generation viable, Hy Image3.5 preview is the one with a rate card, and $0.024 per image is a real number you can put in a spreadsheet today.

The uncomfortable part of the comparison is that neither claim is independently verified against the other. Muse Image has an arena placement from strangers; Hy Image3.5 has a designer study from its own maker. Neither vendor has published a head-to-head, and no third party has run one. If the choice matters to your business, the only honest way to make it is to run your own prompt set through both — and to treat every performance figure in this article, on both sides, as a claim rather than a measurement.

OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor price change is live the same day.