Good morning. It’s Friday, April 10th.
OpenAI’s Sora video generator went offline last week just in time for ByteDance's flagship video generation model Seedance 2 to flood the market today. Their API was restricted from the US due to extreme copyright violations (it was pressured with legal actions from big Hollywood studios), but 3rd party video generation websites like Lovart and Runway made it available to US customers today.
Seedance 2 is a clear step above the competition, and excels at character consistency between shots. Expect to see your social feeds flooded with these videos.
Check out this short 3-minute film Dragon Blue on Youtube that was made entirely with the model. If you’re interested in making film, there’s never been a better time to start!
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
In partnership with Depot
CI That Actually Keeps Up
Most teams ship faster than their CI can handle.
Depot removes the wait with instant job startup, parallel execution, and real debugging tools so you can move at the speed you build.
– Jobs start in seconds
– Parallel steps inside one job
– SSH into live runs + replay failures
5 min → ~35 sec.
Thank you for supporting our sponsors!
Claude gains ‘Managed Agents’ + smart ‘Advisor’ layer
Anthropic is pushing Claude into a more full-stack agent platform, rolling out Managed Agents that split execution into a cloud orchestration layer with separate “brain,” tool sandboxing, and persistent session state. The setup will have faster production deployment of multi-agent workflows.
Claude Cowork is now generally available on macOS and Windows with SCIM-based role controls, per-team budgets, usage analytics, and OpenTelemetry tracing that feeds into SIEM tools. It also picks up Zoom integration and tighter connector permissions.
A new Monitor tool lets background scripts wake the agent only on specific triggers like log errors or CI/PR updates, replacing constant polling with event-driven execution to reduce token usage in long-running workflows. A new ‘advisor’ tool pairs cheaper Sonnet or Haiku executors with Opus as an on-demand reasoning layer that steps in only for hard decisions, improving performance while cutting per-task costs through selective escalation.
Anthropic is also exploring custom AI chips as compute demand spikes, even as its “Mythos” security model raises SaaS disruption fears after showing strong vulnerability-finding and coding capabilities. Read more.
Google turns Gemini into an interactive compute layer with live 3D simulations
Google is turning Gemini into something closer to an interactive compute layer. The app can now generate simulations and 3D models in chat that update in real time as you adjust variables like gravity, velocity, or molecular parameters, replacing static diagrams with manipulable systems triggered by prompts like “show me” or “visualize.”
It’s also adding Notebooks in Gemini, which package chats, files, and instructions into persistent projects synced with NotebookLM. In Colab, Gemini gains Learn Mode for step-by-step coding guidance and persistent Custom Instructions, while Google Research’s PaperOrchestra uses five agents to turn experimental logs into LaTeX papers with near-human citations.
Outside the app, YouTube Shorts is rolling out AI avatars built from recorded face and voice data, with usage limits for creators. Google is also expanding Xeon and IPU work with Intel to support inference-heavy, agent-style workloads. Read more.
OpenAI drops Pro price to $100, clarifies gated cyber product, and eyes retail IPO shares
OpenAI just launched a $100/month ChatGPT Pro tier for heavy Codex users, delivering roughly 5x the usage limits of the $20 Plus plan for agentic coding, cloud tasks, and code reviews, with GPT-5.4 access and rolling windows. The $200 Pro tier stays for even higher volumes, while Plus gets rebalanced for shorter sessions. It undercuts rival Pro plans from Anthropic and Google.
OpenAI clarified an earlier Axios report that its invite-only “Trusted Access for Cyber” as a standalone defensive product, extending the GPT-5.3-Codex pilot, with API credits for vetted partners doing vulnerability discovery and code analysis, echoing Anthropic’s cautious Mythos rollout over dual-use risks.
Valued at $852 billion, OpenAI will reserve IPO shares for retail investors after strong private placement demand, as CFO Sarah Friar said it prepares to act like a public company, possibly as early as Q4. Read more.
Meta rolls out ‘Muse Spark’, its first superintelligence model
Meta just dropped Muse Spark, the first model from its Superintelligence Labs after Llama 4 fell short. This compact, natively multimodal system brings tool use, visual chain-of-thought, multi-agent orchestration, and a “Contemplating Mode” for parallel reasoning, built with fresh pretraining, reinforcement learning, and thought compression. It scores top-five on the Artificial Analysis Intelligence Index and shines in science, math, health, and creative tasks, though it still trails rivals in coding.
The model is already live in Meta AI and comes with a private API preview plus future open-source plans. Hours after launch, Pliny the Liberator leaked its full system prompt, which reveals a design philosophy emphasizing truth, connection, and nuanced reasoning.
At the same time, Meta is pulling top engineers into a new Applied AI Engineering unit to build agentic tools for autonomous coding and shipping. To fuel the push, the company added a $21 billion GPU cloud deal with CoreWeave (on top of $14.2 billion prior), pushing annual capex toward $135 billion. Shares jumped over 8 percent. Read more.
HeyGen’s Avatar V generates high-resolution, identity-preserving talking avatar videos from a single reference video and audio input.
Brila builds one-page websites from real Google Maps reviews, using AI to extract authentic customer insights and messaging.
Velo turns your screen recordings into ready-to-share video messages with AI scripts, avatars, and simple editing.
NovaVoice lets you dictate, control apps, and get AI assistance, all hands-free, using natural voice commands.
Offsite lets you run teams of humans and AI agents together, coordinating tasks and actions in real time from one dashboard.
Thank you for reading today’s edition.
Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.
Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!
Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.
