For a long time, AI video meant pretty pictures and then a second trip for the soundtrack. That split is closing. Two of the tools that matter most for short, directed clips now spit out picture, dialogue, and effects in one pass — and they sit close enough to each other that a creator should pick with eyes open, not vibes.
This week’s bench is that choice: Pika’s new H3 Flex vs HeyGen Video. Same job class. Different strengths once you’re past the demo.
On the bench
Pika H3 Flex vs HeyGen Video
Why this pair: Late October’s useful news for video people wasn’t another “looks better” claim. It was picture-plus-sound becoming the default for short scene generation. Pika launched H3 Flex on Oct 29 (post-trained on MiniMax H3; open weights; up to 12 references; separate audio stems). HeyGen brought HeyGen Video into Creative Studio on Oct 21 and its October API launch discount ended Oct 31, so November pricing is the real sticker. Both make 5–15 second cinematic clips with synchronized audio from text, image, or references — the clearest like-for-like pair in this week’s news.
Pika H3 Flex
Best for: Creators who iterate hard on short directed shots — product films, motion graphics, multi-shot gags, camera choreography — and want to draft cheap/fast, finish at higher res, then take the mix apart in an NLE.
How it works: You prompt (or feed references) and get a clip with native stereo audio. Modes cover text-to-video, first/last-frame image-to-video, multimodal reference-to-video, and prompt/reference-guided editing. Up to 12 reference assets (images, video, audio). Native generation is 480p/768p; finishing upscales the same take to 1080p, 2K, or 4K. Through the API you can request separate Dialogue / Music / Effect stems as WAVs. Durations are 5–15 seconds. Open weights are on Hugging Face with inference code on GitHub; the hosted upsampler and stem separation are Pika-hosted features.
Strengths
- Fast iteration path: draft at 768p, finish the keep at higher res without regenerating the take
- Up to 12 mixed references and reference-guided edits (swap a prop, keep the performance)
- Exportable audio stems so you can duck music without ducking dialogue
- Open weights for teams that want to self-host or experiment
- App plans also expose other models (Kling 4.0, Seedance, Wan, etc.) in one credit pool
Weaknesses
- Short clip cap (5–15s); longer scenes mean stitching
- “Seven seconds for a 15s clip” is transformer compute in Pika’s benchmark, not end-to-end wall clock
- Physics, multi-scene consistency, and heavy action-reference control are areas Pika itself flags as still improving
- Hosted 1080p–4K is upsampling, not native generation
Price (checked Nov 4, 2026)
App (pika.art/pricing, page last updated Oct 28, 2026): Free $0; Starter $10/mo monthly or $8/mo billed annually ($96/year), 900 credits/mo; Creator $35/mo monthly or $28/mo annually ($336/year), 3,150 credits/mo; Fancy from $95/mo monthly or from $76/mo annually, 8,550+ credits/mo; Enterprise custom. All listed individual plans include all models, no watermark, commercial license. H3 Flex at 768p · 5s costs 10 credits; Creator and Fancy currently include free H3 Flex generations every day (limited-time, per pricing page).
API (pika.art/pika-h3-flex): text/image/keyframe per output-second — Launch $0.01 (480p) / $0.02 (768p) / $0.035 (1080p) / $0.04 (2K) / $0.08 (4K); Regular $0.02 / $0.04 / $0.07 / $0.08 / $0.16. Reference-to-video Launch $0.03–$0.10 and Regular $0.06–$0.20 (bills input video seconds + output). Audio stems $0.001/sec. As of Nov 4 the product page still lists both Launch and Regular columns; confirm which rate your account is billed at before you budget.
How creators use it: Rough the shot at 768p until camera and performance feel right, finish one keep, pull stems into Resolve/Premiere, and keep product or character consistency with reference packs across a set of shorts.
HeyGen Video
Best for: Creators and teams who want scene generation inside a broader video workspace (avatars, voice, translation, templates) — product B-roll, lifestyle clips, and short scenes that sit next to presenter video — with simple per-second API pricing.
How it works: One model does text-to-video, image-to-video, and reference-to-video. Each clip returns picture plus synchronized dialogue, ambience, and effects in the same call. Clips run about 5–15 seconds. Available in HeyGen Creative Studio (model picker) and via API / OpenRouter / Runware / ComfyUI. Free product page offers 2 clips a day for short demos. HeyGen’s own materials say it ranked #1 in their blind evals against several peers; treat that as vendor testing.
Strengths
- Picture + sound in one pass, aimed at everyday business and creator scene work
- Lives next to Avatar, Voice, Video Agent, and translation in the same app — useful if you already ship presenter video there
- Clear API rates; fixed seeds for one-change-at-a-time iteration (per HeyGen’s cost guide)
- Creator plan already includes HeyGen Video 1 (text, image, and reference to video)
Weaknesses
- Same short-clip reality: 5–15 seconds, so longer sequences are assembled
- Closed weights; no audio-stem export like Flex
- 1080p / 2K API tiers cost more (HeyGen’s Oct 26 cost guide: 1080p and 2K bill at 3× the 768p rate)
- October’s 50% API launch promo ended Oct 31 — November budgets should use standard rates
Price (checked Nov 4, 2026)
App (heygen.com/pricing and credit plans help): Free $0 (1–3 watermarked videos/mo depending on region, up to 1 minute each); Creator $29/mo (FAQ also lists ~$24/mo if paid annually), 600 credits, 1080p export, watermark removal, HeyGen Video 1 included; Pro from $49/mo (1,000 credits; scales to higher credit tiers, FAQ up to 100,000 credits/mo), 4K export, videos up to 30 minutes; Business $149/mo for first seat + $20/seat/mo, 1,500 shared credits, videos up to 60 minutes; Enterprise custom. API wallet is separate pay-as-you-go.
API / per-second (heygen.com/models/heygen-video-1, HeyGen cost-per-second guide updated Oct 26, 2026): standard from November — about $0.02/sec at 480p and $0.03/sec at 768p for text/image-to-video with native sound; reference mode higher (guide cites ~$0.06/sec at standard rates for reference at the priced tier); 1080p/2K at 3× the 768p rate per that guide. Model page headline: “$0.02 per second” for API access. Confirm live rates in-product before a big run — the October launch discount is over.
How creators use it: Generate a product or lifestyle beat with sound already attached, drop it beside an avatar walkthrough in Creative Studio, or pipe volume B-roll through the API wallet when you need many short takes without babysitting a second audio tool.
Verdict
If your week is iterate the shot — camera language, multi-shot timing, then a real mix in your editor — start with Pika H3 Flex. Stems and the draft-then-finish path are the practical edge.
If your week is ship scenes inside a video production habit you already run on HeyGen (or you want one vendor for presenters and B-roll), pick HeyGen Video. The workspace around the model is the edge.
Neither replaces a long edit. Both replace the dumb loop of “generate silent clip → hunt for temp music → hope the cut still works.” For short directed work in November 2026, that’s the right problem to solve.
New this week
- Higgsfield Cinema Studio 4.0 (Oct 28): Director’s Panel, up to 30-second multi-shot scenes, and up to 50 references (was 15s / 9 refs); native audio; credit costs shown before generate (e.g. 10s @720p with audio = 70 credits; 30s @1080p = 480 credits, or 360 under a limited 1080p discount).
- VEED Dubbing API (self-serve live Oct 28): translate speech, clone each speaker’s voice, lip-sync; Standard $0.11/sec (100 languages), Pro $0.18/sec (multi-speaker, accents, keep background audio); clips up to 10 minutes.
- Synthesia’s ChatGPT plugin (Nov 3): live on every Synthesia plan — describe a video in ChatGPT, revise in-chat; Syren by default, Studio when you need exact scripts; can trigger from connected workflows.
- Google Vids custom avatars from an uploaded photo (announced Nov 2): create a project avatar from Drive/Photos/local image — rollout paused Nov 3 until Google posts a resume date.
- StoryCut’s November update (Nov 4): YouTube auto-import to clips, 18 edit styles, full-frame stock/AI B-roll, filler/retake cleanup, and headline looks that avoid covering the speaker’s eyes.
How this was made
AI assisted research and a first draft of this Wrench Report. A human reviewed, edited, and is accountable for the judgment and the final take.
