In partnership with
Welcome, AI enthusiasts
The knock on AI video has always been the same, which is that nothing holds together long enough to tell a story. ByteDance took Seedance 2.5 global today and one unbroken take now runs three full minutes, with the sound written at the same time as the picture. Studios were already sending legal letters over the version that capped out at fifteen seconds. Let's dive in!
In today’s insights:
Hollywood's Chinese AI Problem Just Got 10x Worse
Anthropic's Mythos 5 Hacked a Security Firm
GPT 5.6 Got 80% Cheaper After Optimizing Itself
Read time: 5 minutes
LATEST DEVELOPMENTS
Evolving AI: Seedance 2.5 opened worldwide today and its long mode runs twelve times the old ceiling.
Key Points:
Standard generation runs 30 seconds in one continuous pass and a long video mode carries a single take to 180 seconds against the 15 second ceiling on Seedance 2.0.
A single generation now takes as many as 50 references including green-screen plates and untextured 3D white models.
Plugins for Maya and Blender feed scenes straight out of a working film pipeline into the model.
Details:
ByteDance launched Seedance 2.5 on Dreamina today for paying subscribers across every market it serves. The model writes sound and picture at the same time so dialogue and background audio come out matched to the footage. Editors can redo one part of a frame and leave the rest of the shot alone while timing is set to the exact second. Camera moves and performance blocking are directed the same way a crew would handle them on a real set.
Why It Matters:
In February a user pointed Seedance 2.0 at Tom Cruise and Brad Pitt, and the clip drew the first cease and desist the Motion Picture Association had ever sent an AI company. ByteDance paused the worldwide rollout and came back with filters that refuse real faces. Those filters are still on, and Seedance 2.5 came out stronger regardless, because blocking a famous face does nothing about how convincing the footage around it looks.
TOGETHER WITH DATADOG
🔐 LLM Observability Best Practices
Evolving AI: 4 Key Insights for Scaling LLM Applications.
LLM workflows can be complex, opaque, and difficult to secure. Get the latest ebook from Datadog for practical strategies to monitor, troubleshoot, and protect your LLM applications in production. You’ll get key insights into how to overcome the challenges of deploying LLMs securely and at scale, from debugging multi-step workflows to detecting prompt injection attacks.
RED TEAM FAILURE
🚨 Anthropic's Mythos 5 Hacked a Security Firm
Evolving AI: Claude built working malware for a fake target, and it went live on the public internet instead.
Key Points:
Mythos 5 was one of three Claude models that ended up on live systems, and it never targeted the security firm or knew the firm existed.
The model had already worked out that publishing would count as an attack, then talked itself back into believing the environment was staged.
The firm's scanner installs public packages as routine, so it ran the hidden code, and the credentials it gave up let Claude reach deeper infrastructure.
Details:
Anthropic ran the exercise inside a test range built by an outside partner, Irregular, where a wrong setting left the machines wired to the open internet while the prompt said otherwise. All three cases surfaced only after OpenAI disclosed a similar breakout, which prompted a sweep of 141,006 test runs. The package Mythos 5 published reached 15 machines in the hour before Python's registry pulled it. Opus 4.7 worked out that its own target was a genuine company and took hundreds of rows from a live database anyway, while an unreleased internal model reached that same conclusion and shut its run down.
Why It Matters:
Claude's own reading of its situation is the only thing that ended that third run, which means a security judgement now sits inside the model itself. Anyone who installs public packages or keeps a service reachable from the internet is downstream of that judgement, and two of the three models here got it wrong.
Privacy-first email. Built for real protection.
End-to-end encrypted, ad-free, and open-source. Proton Mail protects your inbox with zero data tracking.
AI PRICE CUT
💸 GPT 5.6 Got 80% Cheaper After Optimizing Itself
Evolving AI: GPT 5.6 Sol rewrote the code that runs OpenAI's models, and Luna's price fell 80% a day later.
Key Points:
Luna's rate falls to $0.20 per million input tokens and $1.20 per million output, with Terra 20% cheaper and Sol offered an optional fast lane at twice its usual price.
Sol's rewrite touched the serving layer alone, leaving the model's weights and its intelligence untouched.
Auto-review in the ChatGPT app and Codex CLI now runs on Luna instead of GPT 5.4, which OpenAI expects to cost about 10x less.
Details:
Codex is where the work happened, and Sol began by studying live production traffic to find imbalances engineers had been approximating by hand. It then rewrote the GPU kernels that run the model's math, cutting end-to-end serving costs by 20%. Hundreds of its own experiments on the small helper model that drafts tokens ahead of it lifted output efficiency past 15%, and OpenAI verified the new code with an open-source checker before shipping.
Why It Matters:
ChatGPT and Codex users receive that shipped code as a plan that stretches further, since one subscription now covers more work before limits arrive and jobs nobody would have paid to automate become worth handing off. The savings arrived because the system audited itself, and that loop is set to run again.
👀 Click on the image you think is real
QUICK HITS
🛡️ Onyx Security Raises $113M Series B at $640M Valuation to Control AI Agents.
📊 groundcover Raises $100M Series C to Build Observability for the AI Era.
💰 Encore AI Raises $30M Series A to Scale Revenue-Focused Enterprise AI Agents.
🛡️ Mate Security Raises $35M Series A to Build an "Operating System" for AI Cybersecurity Agents.
📈 Trending AI Tools
📝 Granola - AI notetaker that captures the real insights and turns every conversation into ready-to-share, action-driving notes*.
🤖 Auto Learning Agents - Open-source self-hosted AI agents that learn from every task.
🛋️ MeltFlex - Redesign any room with real shoppable furniture in seconds.
🌀 Depthflow - Turn flat 2D images into 3D parallax motion.
*partner link