Wan 2.5: the complete guide

Wan 2.5 is part of Alibaba's open-weights-influenced video family — strong in the value tier for social clips, fantasy and action scenes, with bilingual (Chinese and English) prompt support. It is not a native speech/lip-sync model: audio enters through an optional background music URL, not generated dialogue.

• ▸Per-second billing on fal scales linearly with resolution — 1080p is 3× the 480p rate.
• ▸Prompt expansion (default on): an LLM rewrites short prompts; disable when you need exact copy.
• ▸Safety checker enabled by default — factor moderation failures into retry budgets.
• ▸Wan 2.6 successor adds multi-shot prompting, 15 s duration and more aspect ratios — check endpoint before new projects.
• ▸Open-weights heritage: self-hosting paths exist for teams that want zero marginal per-clip cost at the expense of GPU ops.

Official fal.ai rates for Wan 2.5 preview text-to-video:

• ▸Background audio via audio_url does not change the published per-second video rate on fal — you supply hosting for the MP3/WAV.
• ▸10-second 1080p at $1.50 sits between Hailuo Pro ($0.48 max 6 s) and Seedance fast ($2.42 @ 10 s 720p) — compare usable quality, not sticker price alone.
• ▸Processing time: fal docs cite 1–3 minutes typical; longer for 10 s 1080p.

Primary endpoint: fal-ai/wan-25-preview/text-to-video

curl -X POST https://queue.fal.run/fal-ai/wan-25-preview/text-to-video \
  -H "Authorization: Key $FAL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A white dragon warrior stands on a cliff at dawn, cape fluttering in wind, cinematic wide shot.",
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "duration": "5",
    "enable_prompt_expansion": true
  }'

• ▸Bilingual prompts work — Alibaba's docs explicitly support Chinese and English; pick one language per prompt for consistency.
• ▸Fantasy, action and costume scenes are common showcase categories in Wan examples — lean into clear subject + environment + camera.
• ▸Use negative_prompt for recurring artifacts (extra limbs, blurry faces) instead of long "no X" lists in the main prompt.
• ▸audio_url: host a royalty-free loop; if audio is shorter than video, the remainder is silent per API docs.
• ▸Disable prompt expansion when testing identical-prompt battles — expansion changes wording between runs.
• ▸For square social: 1:1 aspect ratio is supported natively — no crop-from-9:16 workaround needed.

• ▸Wins: flexible aspect ratios at low cost, bilingual prompts, optional scored BGM without a separate edit pass.
• ▸Loses to Hailuo on raw per-second floor at 768p for silent clips.
• ▸Loses to Veo and Seedance on speech-led UGC and reference-heavy brand work.
• ▸LTX-2 wins extreme draft speed and self-host economics; Wan wins hosted API simplicity.

Documentation verdict: Wan 2.5 is a strong value-tier hosted API — especially when you need 1:1 or 9:16 with optional background music. Pair with Hailuo for silent motion tests and escalate to Kling or Veo when architecture polish or speech matters.