Runway Launches Aleph 2.0 and a New Edit Studio App, Bringing Image-Level Precision to Video
If a meaningful share of your video work involves remixing or fixing footage rather than generating from scratch, this is the more important release of the week. Aleph 2.0 brings image-level precision to video editing: hand the model a single still showing the change you want and it propagates that edit through the clip, including across multiple cuts. Up to 30 seconds at 1080p is now in scope, which finally covers most ad and short-form social formats.
The strategic point is that Runway is leaning into edit-then-ship rather than text-to-video. For marketing and post-production teams already paying for Runway, Edit Studio collapses tasks that previously required Premiere or After Effects round-trips: things like product colourway swaps, background changes and seasonal variants. It is available today on all paid Runway plans through the desktop web app, so there is nothing new to install if you already have a seat.
Spotify Wires ElevenLabs into Spotify for Authors for Self-Published Audiobook Creation
This matters because it is the first time ElevenLabs voice generation has been embedded directly into a major distribution platform without an exclusivity clause attached. If you write books and want them as audio without paying for studio narration, you can now run the whole pipeline inside Spotify for Authors and still publish the resulting audiobook anywhere else. The beta opens in June, invite-only, English first, and Spotify confirmed authors are not bound to Spotify-exclusive distribution.
For anyone tracking voice AI commercially, the wider lesson is that ElevenLabs has now landed inside the major audio platform alongside its standalone audiobook product. Combined with the recent investment round, it is a signal that text-to-speech is moving from API curiosity to default infrastructure for new audio. Worth watching for sound creators evaluating which TTS stack to standardise on.
Microsoft and EY Commit $1 Billion to Help Enterprise Clients Move Copilot Pilots into Production
On its own, a billion-dollar partnership sounds like a press release, but the more useful read is what it tells you about where Microsoft thinks Copilot is stuck. The money is earmarked for moving enterprise clients out of pilot purgatory and into production AI projects, with EY consultants doing the implementation work. If you sit in an organisation that has been running a single-team Copilot proof of concept for nine months, expect a renewed push from your account team to scale it.
There is no new product here, but the funding mechanic effectively subsidises the change-management cost that has been the real blocker to Copilot rollouts. For competitors selling into the same buyers (including ChatGPT Enterprise, Claude and Glean), this raises the implicit cost of switching by making Microsoft consulting attention cheaper.
Cerebras Runs Kimi K2.6 at Around 1,000 Tokens per Second, Opening Enterprise Trials
If you build agents on open-weight models, this is the speed milestone to pay attention to. Cerebras is running Moonshot AI's trillion-parameter Kimi K2.6 at around 981 output tokens per second on its custom hardware, independently verified by Artificial Analysis. The practical figure is the example agentic coding request: a 10,000-token prompt plus 500-token completion delivered in 5.6 seconds end-to-end, against 163.7 seconds on Moonshot's own endpoint.
For anyone evaluating where to host larger open models for latency-sensitive workloads such as IDE agents and voice-driven workflows, Cerebras now offers a credible alternative to GPU cloud providers for K2.6 in particular. Enterprise trials are open today.
Corti Launches Symphony for Speech-to-Text, a Clinical-Grade ASR Family Hitting 2% Word Error Rate
Healthcare is one of the few verticals where domain-specific AI consistently beats general-purpose tools, and Corti's new Symphony family is a useful illustration. The models hit a 2 per cent word error rate in clinical environments and outperform general-purpose ASR on medication names, dosages and clinical shorthand. Premier-tier models are shipping for English, German, French and Danish, each validated against more than 150,000 medical terms. If you are building any kind of clinical scribe, telemedicine tool or documentation workflow, this is the kind of specialised API worth swapping in for the default Whisper or Google Speech endpoints.
Industry Themes
Three patterns stand out today. First, the product axis has shifted away from the frontier labs and onto the application layer: the most concrete launches were Runway's video editor and Spotify's ElevenLabs-powered audiobook tools, both surfaced through the companies' own channels before the press picked them up. Second, infrastructure for open models is now a competitive front in its own right, with Cerebras using Kimi K2.6 to make the case that custom silicon is no longer a niche play. Third, vertical specialisation is winning where general models are good enough but not great, illustrated by Corti's clinical ASR beating Whisper on medical terms. The picks-and-shovels story is more interesting than the model story this week.