Synthesia puts video generation inside ChatGPT as new models settle on 30 seconds

Synthesia moves production into the chat window

Buyers now pick a pipe rather than a model: a chat plugin, an API, or an aggregator that resells a dozen models at once. That shifts price comparison onto credit packs and per-seat plans. Watermark and commercial-use terms become the deciding detail for small teams.

The plugin does more than return a clip. Users describe the video or hand over a document, deck or spreadsheet, and ChatGPT asks follow-up questions about style and length before Synthesia builds it. Further requests in the same conversation change the result. Videos use Syren by default, which Synthesia pitches at explainers, how-tos and launch videos. A prompt can ask for Studio instead, where an avatar delivers a supplied script word for word. The result opens in Synthesia's editor for scene-by-scene changes.

The automation examples are the interesting part. Synthesia says a workflow can turn each new post in a product-releases Slack channel into a 30-second launch video. Another checks BambooHR at 8 a.m. on weekdays and makes a 60-second welcome video for each new hire.

That is a different sales pitch from a model benchmark. The video is an output of a process the team already runs.

The new models agree on 30 seconds and sound

Kling 4.0 is live on fal's playground and fal Agent, according to a DotAI Times write-up on note.com. It returns 3 to 30 seconds of video with two-channel stereo audio in one generation. Dialogue, singing, sound effects and ambience are generated in sync with the picture. One run can reference up to 15 assets: 10 images, 5 videos totalling 30 seconds, and 7 elements.

ByteDance's Seedance line goes further on references. TechFlow reported on 28 October that B.AI added four Seedance models to its API. Seedance 2.5 handles single runs of up to 30 seconds at up to 1080p 10-bit, with up to 50 reference assets and timestamp-targeted editing. Seedance 2.0 targets 4K 10-bit with start and end frame control. Seedance 2.0 Fast generates at 480p or 720p, and 2.0 Mini sits at the bottom on unit price.

So the two flagship releases land on the same duration. Fifteen reference assets against 50 is the wider gap. It matters for anyone building a recurring character or a product line rather than a one-off clip.

Distribution is doing more work than the model

AIVideoFree relaunched on 30 October with sign-up credits that work across all 12 of its video models and 10 image models, including Veo, Kling 3.0, Seedance 2.5, Wan 3.0 and Minimax H3. The company says videos download without a watermark, credits do not expire, and free-credit output can be used commercially. No credit card is needed to sign up, and the company says a subscription is not required.

Cuty AI takes the same idea further. The Jerusalem Post's roundup of team-oriented video tools lists 28 video models on its site. They include Seedance 2.5, Veo 3.1, Kling 3.0 and Wan 3.0.

B.AI sells the other end of the pipe: API access to four Seedance tiers, from the flagship 2.5 down to 2.0 Mini. A team that wants the newest ByteDance model can reach it through B.AI's API, through an aggregator, or through a chat plugin. Each route carries its own terms, its own credit rate and its own rules on commercial use.

What access costs

The listed prices are for credits and seats, not for a model. AIVideoFree sells one-time packs starting at $6.90 for 600 credits, with annual plans from $7.40 per month billed yearly. Cuty AI's yearly-billed Lite plan is $4.9 per month for 300 credits, and Pro is $8.9 per month for 1,000. Adobe Firefly team plans start at $19.99 per licence per month, according to the Jerusalem Post's comparison.

Read those side by side and the gap is $4.9 against $19.99. The comparison is not clean. AIVideoFree shows the credit cost of each generation before it starts. The Jerusalem Post's review also notes that videos made on Cuty AI's free version may carry a watermark, which undercuts the free tier for client work.

For a small team, the deciding question is often not quality. It is whether the output can go to a client without a watermark and without a per-seat contract.

What is still open

Kling 4.0's API access is listed as coming soon, so builders cannot yet wire it into their own products. Video extension up to 2 minutes is promised without a date.

The 28 October Seedance announcement listed capabilities but no per-call prices. Neither B.AI nor ByteDance has published what a 30-second Seedance 2.5 run costs through the API.

That gap is the one to watch. Duration, audio and reference counts have converged fast across Kling and Seedance. Pricing per finished second has not.