Tencent Hy Image 3.5 Preview ($0.024/2K): a fourth routing option for Bubble bulk catalogs in BIAIG

Why now

The digest flagged Tencent Hy Image 3.5 Preview at US$0.024 per 2K output image on international TokenHub—alongside GPT Image 2.5, Gemini Nano Banana 2, and gpt-image-2 batch discounts. For Bubble bulk catalogs routed through BIAIG, this is a fourth candidate, not a wholesale swap. Same topic: gpt-image-2 batch discount: unit economics. A closer look: Batch Banana FLUX 3 Image. Same topic: gpt-image-2 batch discount: unit economics.

Scout-bound facts: release 2026-09-22; model hy-image-v3.5-preview; sync v35-generation with a Messages API (not Hy 3.0’s single-prompt shape); up to 20 reference images (≤20 MB each); 1K / 1.5K (default) / 2K via generate_max_pixels, explicit size up to 4096×4096 for 4K; optional session and assembled_history for multi-turn edits; result URLs expire in ~12 hours. Launch coverage (CNR / Tencent messaging) cites ¥0.15 per 2K on mainland and US$0.024 per image internationally, billing only on successful outputs (refs and failures described as not charged). Official TokenHub lists Hy-Image-3.0 at US$0.032/image, but no Hy-Image-3.5-preview row appeared in the pricing table reviewed here—so $0.024/2K is UNCERTAIN until the rate card catches up.

Unit economics vs sibling routes

Use sibling numbers for orientation only—APIs, resolutions, and input charges differ:

Route	~2K-class output anchor	Notes
Hy 3.5 Preview	~$0.024/output (launch)	Sync 2K; preview lifecycle
Gemini 3.1 Flash Image (Nano Banana 2)	~$0.101 standard / ~$0.050 Batch (Google pricing)	Different contract, batch latency
gpt-image-2	Batch output $15/M tokens (~50% vs standard); 2K not apples-to-apples with Hy tier	Strong if you already live on OpenAI edits

Takeaway: Hy is the cheapest headline for sync 2K if $0.024 holds. True bulk cost = successful outputs × rate plus each multi-turn bill plus immediate fetch to durable storage—provider URLs are temporary.

When to add Hy 3.5 to the BIAIG router

Consider primary when:

• jobs need 2K masters and sync latency is acceptable;
• ≤20 refs cover your SKU/ad variant set;
• you have your own consistency checks on those refs (vendor +30% vs Hy 3.0 is not an independent commerce benchmark);
• margin matters more than typography/grounding where Gemini or OpenAI already win in your tests.

Keep Gemini Nano Banana 2 or gpt-image-2 (especially Batch) as fallback when:

• copy, grounding, or existing OpenAI-native edit flows matter;
• Batch latency is acceptable for margin;
• Hy returns 422 (safety) or 429 (concurrency)—don’t let one provider stall an entire catalog without a defined retry/fallback policy.

Don’t default everything to Hy: preview models may change upstream model strings; Tencent warns against branching on them—GA behavior and pricing can shift.

Primary / fallback without breaking ref/consistency

1. Pin one provider per SKU or batch job when pipelines depend on fixed ref sets, seeds, or edit chains. Hy Messages + assembled_history does not port to GPT/Gemini request shapes.
2. Do not switch providers mid bulk-run for one variant lineage. If primary fails, use an isolated retry job with re-validation (visual QA, SKU color), or restrict fallback to new SKUs—not half-finished series.
3. Lock resolution per job: set generate_max_pixels / size for true 2K; the 1.5K default saves money but breaks consistency with other 2K routes.
4. Persist every Hy output immediately; pipelines that assume permanent provider URLs fail after ~12 hours.
5. Budget multi-turn edits as multiple billed outputs, not a single generation.

Bubble-oriented checklist

• Dedicated BIAIG route: hy-image-v3.5-preview, v35-generation, messages payload.
• Provider lock keyed by job_id or catalog_batch_id.
• Fallback after N retries on 429/5xx only if QA allows mixed lineage.
• Cost guard: max_outputs_per_SKU and alerts when assembled_history chains grow.
• Open: official 3.5 TokenHub row; international 4K price; your benchmarks vs Nano Banana 2 / gpt-image-2 on product/ad refs.

Honest scope

This is routing methodology, not a Hy sales pitch. Without operator benchmarks on your catalog refs, we cannot claim $0.024 buys enough consistency. Until the rate card matches launch messaging, model pricing mismatch risk in BIAIG margins. A closer look: Flare, Sunburst, and WaveSpeedAI: routing.