Welcome back! Anthropic put its most dangerous model in everyone's hands and buried a meter in the fine print. Google moved its live interpreter into your phone's earpiece. OpenAI taught apps built on it to answer in pictures. And the 30-minute production app from Monday's Atlas gets its full breakdown.
In today's Generative AI Newsletter:
Fable 5 (or Mythos) arrives: How many days do you get before they stop your access?
Live translation: Are we about to stop learning different languages?
OpenAI image search: Why will apps built on OpenAI stop answering in walls of text?
The 30-minute app: How does one engineer ship to production without opening a console?
Yesterday we flagged prediction markets betting Anthropic's exploit-finding Mythos model would go public within days. Hours after we hit send, it happened.
Fable is Latin for what mythos is in Greek. It's live for everyone, and it's the most capable model Anthropic has ever shipped to the public. Claude Fable 5 is the same as Mythos but only different in the safeguards.
The rest of the story sits in the fine print:
The meter: Fable is included on Pro, Max, Team and Enterprise seats through June 22. On June 23 it comes off those plans and runs on usage credits until capacity catches up.
The difference: Ask about cybersecurity, biology or chemistry and Claude Opus 4.8 answers instead.
Your data: A new policy comes with it. Anthropic will keep all Mythos-class traffic for 30 days, even business customers, to catch jailbreaks, and says the data won't train new models.
It’s already doing great things. Stripe pointed Fable at a 50-million-line codebase and got a migration done in a day that would have taken a full team over two months by hand.
It needs less guidance than older models as well. Fable beat Pokémon FireRed from raw screenshots, with none of the helper tools earlier Claudes needed.
Until Monday Mythos was the model Washington wanted kept from China, capped at roughly 200 vetted security organizations. Now Fable, built on Mythos tech, is here for $10 per million input tokens, and it’s available on paid Claude plans until June 22.
Special highlight from our network
Anthropic has been shipping new Claude capabilities almost every week. Skills, Connectors, Cowork, vibe coding. Most professionals are still using Claude the way they used ChatGPT a year ago, missing the modes that actually change how work gets done.
Outskill’s 2-Day Claude Mastery Workshop is a live walkthrough of every major Claude capability, taught by mentors who’ve spent 800+ hours testing them in real workflows.
You’ll leave with working setups, not theory: Connectors automating your Notion and files, vibe-coded dashboards, and a clear sense of which Claude mode to use for what.
Free this weekend.
Speaking of Stripe doing months of work in days, that change has reached everything else too, yet most people are still using AI as a Google Search alternative.
Now you can build a global startup and VC intelligence platform from scratch with AI, and Nawas decided to show us how.
Nawas Naziru Adam, an AI and robotics engineer who has led teams putting AI into embedded systems and teaches LLMs and data science on Udemy, runs a 4-week cohort course where you build a system from scratch in 4 weeks using AI.
It walks through:
Cleaning and merging messy multi-source startup data with generative AI
Running network analysis on funding patterns
Shipping the insights as an interactive dashboard
Building your version from scratch so you get actual, practical experience
You learn by building your project. So you leave with a working portfolio piece and a process you can point at any job. Save your seat here.
Hold your phone to your ear like a regular call and the person talking comes through in your language, a few seconds behind, with their tone and pacing intact. No headphones needed.
That's listening mode, new on Android, and it ships with Gemini 3.5 Live Translate, the new audio model now behind live translation in the free Google Translate app across 70+ languages.
The model is why conversations stop stalling. It translates continuously while you speak instead of waiting for you to finish, so the audio never pauses for its turn. The headphone version of live translate has been in the app as a beta since December. Yesterday Google swapped in the new model and took it global, on Android and iOS.
The rollout covers all three audiences at once.
Developers get the model in public preview through the Gemini Live API. Google Meet picks it up this month in a private preview for select business customers, jumping from five languages routed through English to 70+ in any direction, over 2,000 combinations in one meeting.
Grab, Southeast Asia's ride-hailing giant, is testing it on driver pickups, where users make over 10 million voice calls a month.
A six-month beta just became a global default. Next time a call, a meeting or a taxi ride lands in a language you don't speak, the fix is already on your phone.
OpenAI added image results to the web search tool in its Responses API. An app built on it can now pull product shots, places and visual references straight from a search, each with a caption and a link back to the page it came from.
For builders it's one setting. Add image to search_content_types, cap how many results you want and the tool hands back URLs, thumbnails and captions ready to render.
The deadline is real too. OpenAI's older search-preview models shut down on July 23, so anyone on the legacy path has a reason to migrate and something new to build when they land.
If your product answers questions with cited paragraphs, it can start answering some of them with pictures. Shopping, travel and research apps get the obvious win.
Monday's Technical Atlas covered a Google AI engineer who built and deployed a production application in about 30 minutes with Claude Code, work that usually takes five people and weeks of handoffs.
He did it by running every traditional software stage at the same time, and we promised the full breakdown.
Here it is.
1- Plan Mode wrote the spec before any code. Claude drafts the full specification first. You read it, edit it and approve it before a line gets written. That stops the "just build it" followed by a bug-fixing spiral that eats your day.
2- Subagents ran the workstreams in parallel. One agent per workstream, API, data pipeline, dashboard, each one assignable to a different model. A real team sprint but fully simulated.
3- MCP servers kept the docs current. Google's Developer Knowledge API refreshes every 24 hours, so Claude works from documentation that's never stale and picks the right products for the job, Cloud Run, Firestore, BigQuery.
4- Security plugins reviewed while building. You codify your own rules, input validation, IAM least-privilege, the OWASP top 10, the standard list of web security risks. The plugin enforces them before anything moves from dev to prod. The plugin is the security review.
5- Skills handled the deployment. Official Google Cloud skills run CI/CD end to end, Cloud Build for integration, Cloud Deploy to promote dev to staging to prod. You never open the console.
Five phases used to mean five teams and a delay at every seam.
Here they read like one long conversation, and the bottleneck moves somewhere new, deciding what a team even looks like when one person can ship the whole stack.
Jina AI turns any web page into clean text an AI can actually read. Its Reader strips a URL down to pure markdown, with no menus or cookie banners, which is the format models digest best. It runs straight from the browser with no account, and it's the plumbing a lot of AI apps already run underneath.
Try this yourself:
Put r.jina.ai/ in front of any article URL in your browser and the page comes back as clean markdown.
Paste that output into ChatGPT or Claude when a chatbot can't open a link you want it to read.
Swap r for s, as in s.jina.ai/your+search+terms, to get live web search results in the same clean format.
Building something? Grab an API key at jina.ai to raise the rate limits and plug Reader into your agent.
Who it's for: anyone feeding web pages to an AI.
Meta has five days to open WhatsApp: The EU ordered Meta to give rival AI chatbots free WhatsApp access within five working days or risk a fine of up to 10 percent of its turnover, after rejecting Meta's pay-for-access offer in April. Meta says it will appeal.
Gas will feed the AI buildout: US Energy Secretary Chris Wright said natural gas will be the biggest new source of electricity to power AI, "the thing we can grow rapidly for 24/7 reliable firm electricity today."
Microsoft's repos turned into traps: Microsoft disabled at least 70 of its GitHub repositories after hackers injected malware that steals a developer's credentials when the code is opened in AI coding apps. It's the second time in weeks hackers got into Microsoft's open source projects.
Outside economists get OpenAI's data: OpenAI opened the Economic Research Exchange, giving selected researchers privacy-protected access to its tools and datasets to study what AI is doing to workers, firms and the broader economy. Proposals close July 5.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
🐝 Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
