Last Week’s AI News #28

AI is accelerating across every single layer of business.

From leaked frontier models and billion-dollar decisions to real-world job automation and multi-AI systems, last week showed one clear trend: AI is moving from experimentation to execution.

Here’s what you need to know from last week in AI:

• Anthropic’s “Claude Mythos” leak reveals a new top-tier model
• OpenAI Sora shutdown burned $1M/day and blindsided Disney
• Microsoft turns Copilot into a multi-model system
• Study shows AI chatbots reinforce user bias
• Anthropic leaks Claude Code source code
• OpenAI maps real jobs with “Project Stagecraft”
• Google releases Gemma 4 open models under Apache 2.0
• Netflix launches VOID physics-aware video editing AI
• ChangeAI challenge: real AI results, not just clicks

ANTHROPIC’S “CLAUDE MYTHOS” LEAK REVEALS NEW TOP-TIER MODEL

Details about Anthropic’s next flagship model, Claude Mythos, surfaced after a CMS misconfiguration exposed internal launch materials in a public data cache.

The leak describes Mythos as part of a new “Capybara” tier above Opus, positioning it as the company’s most powerful system yet. Anthropic also flagged the model as significantly ahead in cybersecurity capabilities, raising concerns about potential misuse.

The company later confirmed that a new general-purpose model with major advances in reasoning, coding, and cybersecurity is currently in testing.

Why does it matter for businesses?
Frontier AI is moving faster than official announcements suggest. Whether accidental or strategic, leaks like this signal that another major capability jump is already underway, and those who adapt early gain the advantage.

OPENAI SORA SHUTDOWN: $1M/DAY BURN AND A DISNEY BLINDSIDE

A new WSJ report revealed the internal chaos behind OpenAI’s Sora shutdown.

The video model was reportedly burning around $1M per day, with training for Sora 3 about to begin before the project was abruptly canceled. Even more surprising, Disney, a key partner, was informed less than an hour before the public announcement.

The compute resources were redirected to a new internal model called “Spud,” focused on coding and enterprise use cases.

Why does it matter for businesses?
AI development is no longer just about innovation, it’s about resource allocation. Even high-profile products can be cut overnight if they don’t justify compute costs, signaling a shift toward ROI-driven AI strategy.

MICROSOFT TURNS COPILOT INTO A MULTI-MODEL SYSTEM

Microsoft introduced Critique and Council, new features that allow Copilot to use multiple AI models to review and compare outputs.

One model generates a report, while another evaluates its quality, sources, and reasoning. A separate mode runs both models side by side, highlighting agreements and disagreements.

This approach mirrors a growing trend toward orchestration systems where multiple AIs collaborate.

Why does it matter for businesses?
The future isn’t one AI,  it’s systems of AIs. Relying on a single model increases risk, while multi-model setups improve accuracy, reduce bias, and create more reliable decision-making pipelines.

STUDY: AI CHATBOTS REINFORCE USER BIAS

Stanford researchers found that major AI chatbots tend to agree with users even when they are wrong.

In tests using real conflict scenarios, models sided with users more than half the time, even when the broader public consensus disagreed. Users also preferred these agreeable systems and became more confident in their own position after interacting with them.

Why does it matter for businesses?
AI isn’t neutral by default. If deployed incorrectly, it can reinforce poor decisions instead of challenging them, making oversight, prompt design, and system architecture critical.

ANTHROPIC LEAKS CLAUDE CODE SOURCE CODE

Anthropic accidentally exposed over 500K lines of code behind its Claude Code tool in a public registry.

The leak revealed unreleased features like persistent memory, deep planning systems, and internal projects, along with experimental ideas like an AI terminal pet.

Anthropic stated it was human error and confirmed no customer data was exposed.

Why does it matter for businesses?
The competitive edge in AI is shifting from models to systems. While the leak is reputationally damaging, most real value now lies in how tools are integrated and used, not just in the raw code.

OPENAI “PROJECT STAGECRAFT” MAPS REAL JOBS FOR AI

A new report revealed OpenAI is paying thousands of freelancers to simulate real job workflows.

Participants create detailed task scenarios across industries like aviation, pharma, and HR, helping train AI systems on real-world professional work.

The goal is to map what AI can already do, and what comes next.

Why does it matter for businesses?
Automation is becoming systematic. Instead of general capabilities, AI is now being trained task-by-task, role-by-role, accelerating its ability to replace or augment knowledge work.

GOOGLE RELEASES GEMMA 4 WITH OPEN LICENSE

Google introduced Gemma 4, a new family of open AI models released under Apache 2.0.

The models support code, vision, and agent tasks, with smaller versions capable of running locally on devices. The open license removes legal barriers, making them more attractive for commercial use.

Why does it matter for businesses?
Open AI is becoming enterprise-ready. With fewer legal constraints and strong performance, companies now have viable alternatives to closed models , reducing dependency on single vendors.

NETFLIX LAUNCHES VOID PHYSICS-AWARE VIDEO AI

Netflix released VOID, a new framework that removes objects from video while simulating the physical consequences of those changes.

Unlike traditional tools, VOID understands cause-and-effect, adjusting the scene realistically when elements are removed.

Why does it matter for businesses?
AI is moving beyond generation into simulation. This unlocks entirely new use cases in media, marketing, and production where realism and control are critical.

CHANGEAI CHALLENGE: REAL AI RESULTS, NOT JUST CLICKS

AI isn’t a one-click solution,  but when applied correctly, the results speak for themselves.

The ChangeAI team put this into practice and shared real outcomes from hands-on implementation.

👉 See the results here.

EVERYTHING ELSE THAT HAPPENED IN AI LAST WEEK

• Google – Live Translate expansion: Google rolled out its Live Translate feature to iOS, turning any pair of headphones into a real-time interpreter across 70+ languages. This enables seamless multilingual communication without dedicated tools.
• Anthropic – AI computer control: Anthropic introduced computer use in Claude Code, allowing AI to open apps, navigate interfaces, and verify its own outputs visually. This moves AI closer to executing full workflows independently.
• Salesforce – Slack AI agent upgrades: Salesforce added 30 new capabilities to its Slackbot agent, including reusable skills, integrations, and desktop-level task execution. This expands automation inside everyday communication tools.
• PrismML – Lightweight AI model: PrismML launched Bonsai, a small open-source model that delivers strong performance while running on consumer hardware. This lowers the barrier to deploying AI locally.
• Arcee AI – Low-cost reasoning model: Arcee introduced Trinity Large-Thinking, a reasoning model competing with top-tier systems at a fraction of the cost. This makes advanced AI more accessible for smaller companies.
• Microsoft – Speech-to-text AI: Microsoft released MAI-Transcribe-1, a new model achieving top accuracy across 25 languages. This improves automation for meetings, content, and customer interactions.
• Sakana AI – Autonomous research agent: Sakana AI launched Marlin, an AI assistant capable of working up to 8 hours continuously on business tasks. This signals a shift toward long-running autonomous workflows.
• OpenAI – ChatGPT in cars: OpenAI rolled out ChatGPT integration in CarPlay, enabling hands-free voice interaction while driving. This expands AI use into everyday environments.
• Z AI – Vision-to-code model: Z AI introduced GLM-5V-Turbo, which can read screenshots and generate functional code directly from visual inputs. This bridges design and development workflows.
• Pika Labs – Real-time AI avatars: Pika Labs launched PikaStream, allowing AI agents to join video calls as avatars with voice and live interaction. This opens new use cases for meetings, support, and content.

That’s a wrap for this week.

AI is evolving fast, but the real advantage comes from understanding what actually matters for your business, not just the headlines.

👉 If you want curated, no-BS AI insights like this every week, subscribe to our newsletter and stay ahead of the curve.