Changelog

August 7th, 2026

Hedra August 7th Update: Seedance 2.5, the world's most advanced video model, is here

Seedance 2.5 is live on Hedra

  • Seedance 2.5, the world’s most advanced model is here.480p/720p, 4–30s, native audio toggle, reference media up to 50 files. Includes Auto duration: the model picks the length, you're billed the 30s max up front and refunded the unused seconds on completion.

  • Seedance 2.5 is slower and more expensive than Seedance 2.0, it is limited to 720p resolution. Despite this, our testing indicates that users will on average receive better results on the first try with the model.

New models are live on Hedra Create and Develop API (and ½ off).

  • MiniMax H3 — text/image/reference-to-video, now with 4K. We find this to be an excellent daily driver for cost effective generations. We’ve reduced prices by ½ of launch as a promotion on API on the site.

  • FLUX.3 (Black Forest Labs) — text-to-video, image-to-video, and first-last-frame, with native audio.

Model specific prompting skills built into the Hedra Agent

  • Video prompting guidance is now per-model — Seedance 2.0, Seedance 2.5, and MiniMax H3 each get their own prompt authority in the agent, with Seedance 2.5 as the default.

Video generation improvements

  • Duration is a scrubber, not a long dropdown. Seedance 2.5 gets an Auto/Custom toggle.

  • Reference audio and video that exceed model limits get an alert and a per-clip trim instead of a hard failure.

  • Removed the 1.2 MP cap on start-frame uploads; large image attachments no longer blow up memory in the agent.

  • Videos download through the browser's native download instead of an in-app path, making downloads faster and more reliable.

  • Supplying audio to a model with no audio input is rejected up front with a clear reason.

Errors and moderation

  • Backend failure remedies now reach the client as structured reason codes, so error messages say what to do.

  • Model refusals from upstream providers are now clearly communicated to the users.

Visual Skills

  • Assets in skills: a skill body can rewith [[asset:…]], and those assets arereadable by anyone the skill is shared with.

  • Skill editor is now a chip-aware editrence, add assets from the librarypicker, upload, or drag-and-drop; references render as tiles above the instructions and open the asset modal on click.

  • New video extension skill.

  • Personal skills dialog in skill creathidden from the catalog; workflow addedas a skill category.

Agent

  • Edit mode support for Omni Flash, plus an edit card in the agent tool rail.

In app notifications for team joins

  • In-app notification feed — panel, currently fires on workspace invites and canvas shares.

Billing

  • Cancel your subscription in-product instead of being bounced to Stripe. Cancel is hidden if you've already cancelled clearing up some confusion

  • Support-issued credit adjustments now appear in credit history.

  • Fixed: a late-paid old invoice could temporarily lock accounts.

Spaces

  • Upload and render progress pushes live to everyone in the room.

  • A failed node is contained to that instead of taking down the app.

  • Node state is derived from the generation itself, so nodes no longer get stuck on stale room state.New branding

Write a comment...