Runway routes models, FLUX 3 adds audio: July 24

Runway can now choose a media model against your budget, while FLUX 3 can generate video and matching audio in one job. Claude also made spoken work more useful. Give this dispatch 30 minutes and you'll have a capped media test, a cleaner video brief, and a safer voice workflow ready to try.

Runway routes each media job by cost, speed, or quality

Runway launched Media Router on July 23 to choose an image, video, or audio model for each request. A model is the engine that generates the media. The official announcement says you set a price ceiling and decide whether the router should favor cost, output quality, or latency, which simply means how long the result takes to arrive.

This is useful if your business produces the same creative in two stages. Cheap, quick previews can go to one set of models. Approved final exports can use a higher quality setting. Runway says the router can choose among its own models and outside providers, including ElevenLabs for audio. That removes some manual comparison work, but it also means the model behind your output can change without your team choosing it job by job.

Set this up in Runway Dev > Model Router. Create a configuration called social-preview. Choose Cost, enter a hard price cap, and deny any provider your business has not approved for customer material. Turn on Dry run first. A dry run shows which model would be selected without generating media or spending credits. Keep your current fixed-model workflow live until 20 saved prompts produce acceptable choices.

If your media process is still mostly manual, start with the AI avatar video tool setup before adding automatic routing. The router helps a repeated process. It does not create one.

FLUX 3 generates video and matching audio together

FLUX 3 entered early access on July 23 with video clips up to 20 seconds and native audio. Native audio means the same system creates the picture, speech, and sound rather than asking separate tools to assemble them later. Black Forest Labs says in its launch post that the model accepts text, images, or video as references and supports multilingual dialogue, video continuation, and defined start and end frames.

The business consequence is fewer handoffs during early concept work. You can test whether a product action, spoken line, and sound effect belong together before paying for a polished edit. But this is early access, and the maker labels its comparisons preliminary. Keep FLUX 3 off any deadline-dependent production path until access, pricing, and repeatability settle.

Write one reusable brief with five fields: product, action, camera view, spoken line, and required sound. Keep the spoken line under the length that fits naturally inside 20 seconds. Generate three versions from the same brief, then score whether the product stays recognizable and whether the audio matches the action. For finished narration, the line-by-line method in the AI voiceover tool guide gives you more control than accepting one generated soundtrack.

Claude brings stronger models and connected tools into voice mode

Claude voice mode can now use Sonnet, Opus, or Haiku, rather than limiting spoken conversations to the lightest model. Anthropic's current voice-mode documentation also says voice can use connected Gmail, Google Calendar, Google Docs, and Slack accounts. A connected tool is an account you permit Claude to read or use while answering.

That turns voice from dictation into a hands-free briefing layer. You can ask for today's appointments or a summary of a document while preparing for work. It is not a customer-facing phone agent, and it should not become an unchecked route into customer messages. Voice conversations count against normal plan limits, and text transcripts are saved in chat history.

Open Settings > General > Voice settings and select a voice and language. Start a text chat with the model you want, then tap the sound-wave icon. Test with a non-sensitive document first. Ask Claude to summarize it and name the source it used. Review Settings > Connectors afterward and remove any account the test did not need. The same limited-access principle appears in the lead-to-CRM automation runbook: start with sample data before customer records.

On the bench

• Watch for FLUX 3 Image access and public pricing before changing an image budget.
• Test whether Runway's dry-run choices stay stable when you repeat the same brief on different days.
• Compare Claude's voice summary with the source document before using it for a morning briefing.

Use the AI voice and video library for the next controlled media build. Get the next build.