Help make better ads

Did you recently see an ad for Roku Ads Manager in a newsletter? We’re running a short brand lift survey to understand what’s actually breaking through (and what’s not).

It takes about 20 seconds, the questions are super easy, and your feedback directly helps us improve how we show up in the newsletters you read and love.

If you’ve got a few moments, we’d really appreciate your insight.

Take the survey.

Today:

• Anthropic Debuts Claude Fable 5 and Mythos 5

• Apple Previews Next-Gen Siri AI

• OpenAI Confidentially Files for IPO

• Google Launches Gemini 3.5 Live Translate

• Arcee AI Teams Up With Hugging Face

Anthropic Unleashes Claude Fable 5 & Mythos 5

Anthropic just launched its most capable models to date: Claude Fable 5 for general use, and the unrestricted Claude Mythos 5 for cyberdefense and infrastructure. The new models demonstrate state-of-the-art autonomy, handling complex workflows across coding, data analysis, and vision without human intervention.

Key Details:

• Autonomy & Vision: Fable 5 requires drastically less hand-holding than prior models. It can execute months-long codebase migrations in days (as tested by Stripe), and its vision capabilities are so advanced it managed to beat the video game Pokémon FireRed using only raw screenshots—no maps or helper tools needed.

• Mythos 5 for Specialists: While Fable 5 is for the public, Mythos 5 is the exact same model with the safeguards lifted for advanced cybersecurity and biological research. It is being deployed via "Project Glasswing" in collaboration with the US government to secure critical infrastructure.

• Conservative Safeguards: Because Fable 5 is so powerful, Anthropic is playing it safe. If a general user asks a risky question, the system temporarily routes the query to the slightly older Opus 4.8 model to prevent misuse.

• Aggressive Pricing: Both models are priced at $10 per million input tokens and $50 per million output tokens—less than half the cost of the previous Mythos Preview.

Apple’s WWDC26: A Smarter Siri and Unprecedented Parental Controls

Apple previewed iOS 27 and the next evolution of Apple Intelligence, putting a heavily upgraded "Siri AI" at the center of the ecosystem. Alongside system-wide AI upgrades, Apple introduced highly granular parental controls to help families navigate the digital age.

Key Details:

• Siri AI: Siri has been completely rebuilt. It now possesses "on-screen awareness" (meaning it can answer questions about whatever you are currently looking at) and taps into deep personal context across your messages, photos, and emails to get things done.

• App Enhancements: Apple Intelligence brings new features to everyday apps, including "Spatial Reframing" in Photos (allowing you to fix the composition of an image after you've taken it) and "Notify Me" in Safari (which tracks webpages for product restocks or price drops).

• Robust Child Safety: Parents can now set up child accounts with incredibly strict guardrails. Features include "Ask to Browse" for websites, strict app-by-app scheduling, and upgraded Communication Safety tools that automatically block violent or explicit content in Messages and FaceTime.

The Ultimate AI Cash Grab: OpenAI Files for Confidential IPO

The AI arms race is moving to Wall Street. Following on the heels of its chief rival Anthropic, OpenAI has confidentially submitted paperwork to the SEC for an Initial Public Offering (IPO). The move highlights the staggering financial capital required to build the future of artificial intelligence.

Key Details:

• The Filing: Working with Goldman Sachs and Morgan Stanley, Sam Altman's firm is eyeing a potential public debut as early as the fall of 2026. The confidential filing allows the company to keep its financials private while the SEC reviews its prospectus.

• The Infrastructure Race: Training frontier models requires astronomical investment. OpenAI previously told investors it plans to spend roughly $600 billion on AI infrastructure and data centers by 2030, making public markets a necessary step.

• Crowded IPO Window: OpenAI isn't the only giant tapping the market. Anthropic recently filed confidentially as well, and Elon Musk’s SpaceX is currently executing a highly anticipated IPO targeting a valuation of nearly $1.8 trillion.

• Keeping Options Open: OpenAI noted that the timeline is still fluid, stating the filing gives them the option to pull the trigger on an IPO sooner if it makes strategic sense, but they may remain private a bit longer to handle certain internal restructurings.

🧠RESEARCH

SWE-Explore: Benchmarking How Coding Agents Explore Repositories

SWE-Explore is a benchmark, or test, for how AI coding agents—software tools that fix code—find lines before repairing bugs. It checks whether they choose the right files and exact lines, not just whether the final fix works. Across 848 issues, modern agents beat old search tools but still waste context.

Agents' Last Exam

Agents’ Last Exam is a benchmark, or test, for AI agents—software tools that act step-by-step—doing real professional work. Built with 250+ experts, it covers 1,000+ tasks across 55 subfields and 13 industry groups. Results are weak: 2.6% full pass rate, showing agents still struggle with long work and software use.

Latent Spatial Memory for Video World Models

Latent Spatial Memory introduces Mirage, a system that helps AI videos keep objects in the same 3D place across frames. Instead of rebuilding scenes as pixels, it stores memory in “latent space,” a hidden math map. This cuts wasted work, making video creation 10.57x faster and memory use 55x lower.

📲SOCIAL MEDIA

🗞️MORE NEWS

Gemini 3.5 Live Translate Google launched Gemini 3.5 Live Translate, a new audio model capable of translating spoken words across more than 70 languages in near real-time. Instead of waiting for a person to finish talking, it translates continuously while matching the original speaker's tone and speed to prevent awkward pauses. The feature is rolling out to the Google Translate app, enterprise Google Meet, and software developers.

Arcee and Hugging Face Partnership Arcee AI, a rising American artificial intelligence lab, signed a massive deal to use Hugging Face as the exclusive storage platform for all its public and private data, entirely replacing Amazon's cloud storage. This setup allows Arcee to store its data in one central hub while freely moving its processing tasks to whichever computing provider is cheapest on any given day. 

Kimi Work Desktop Agent Kimi Work is a new desktop application for Mac and Windows that acts as an always-on digital employee for office workers. By connecting directly to your computer's files and using an automated web browser, it can work around the clock to pull data, update spreadsheets, or draft reports without your constant supervision. It also features built-in financial data tools and privacy safeguards that ask for your permission before modifying any of your local files.

OpenAI's Master Plan OpenAI has entered the "third phase" of its business, shifting its focus from pure research to making highly capable AI tools affordable and available to everyone. The company plans to build automated AI research assistants by 2028 to speed up scientific progress, with the ultimate goal of giving every person their own highly advanced, personalized AI helper. OpenAI argues that broadly sharing these tools will prevent power from concentrating in the hands of a few tech giants and will heavily boost global economic growth.

Manus Multi-Account Update The AI productivity tool Manus, which was recently acquired by Meta, now allows users to connect multiple Gmail and Google Calendar accounts at the same time. This means the AI can check your work calendar, personal calendar, and client emails all at once to find the perfect meeting time without you having to manually switch tabs. Users can then tell the AI exactly which account to use when sending the final email or calendar invite.

Anthropic's Biology AI Research Anthropic researchers discovered that AI helpers struggle with biology research because scientific databases are messy and built for humans to click through manually. When asked to pull viral data for outbreak research, top AI models frequently made critical errors that would completely ruin a real-world scientific experiment. The team argues that if we want AI to help design drugs or track diseases, the scientific community must rebuild its databases so they can be easily read by automated computer systems.