Sponsored by

Welcome, humans.

AI Independence day just came early, as late yesterday afternoon Anthropic announced Fable 5 is coming back online today.

Technically, that means U.S. export controls on Claude Fable 5 and Mythos 5 were lifted, with Fable 5 returning globally later today on July 1 and Mythos 5 access expanding through approved partners.

That’s a big deal because Fable 5 had become the model people were treating like some forbidden power locked away. Before access disappeared, Matt Shumer used it to build an explorable, screen-accurate 3D Hogwarts castle from one prompt

And now, a dash of cold water: Rob Hallam called it "happy, but mostly disappointed," since routine coding now gets flagged more often and access is capped at half of weekly limits through July 7. Welcome back, have fun, but not for long...

Here’s what happened in AI today:

😺 Anthropic released Claude Sonnet 5 and Claude Science.
📰 Amazon launched a $1B forward-deployed AI engineering org.
📰 Etched exited stealth at a $5B valuation with $1B in signed contracts.
🍪 Google launched Nano Banana 2 Lite and Gemini Omni Flash.
📰 OpenAI introduced GeneBench-Pro, a computational biology benchmark.

… and a whole lot more that you can read about here.

P.S: Want to reach 700,000 AI-hungry readers? Click here to advertise with us.

😼 Claude Sonnet 5 brings Anthropic’s agent push to the default model

Every AI lab wants you to hand more work to agents. The catch: agents get expensive when they need the giant model, and riskier when that model starts touching browsers, terminals, codebases, and company data.

Anthropic’s answer is Claude Sonnet 5, its new default model for Free and Pro users, built to plan, use tools, code, browse, and run longer tasks without needing the pricier Opus tier.

Here’s what happened:

Sonnet 5 is now available across Claude plans, Claude Code, and the API.
Anthropic says it performs close to Opus 4.8 on agentic work, at lower prices.
Intro API pricing is $2 / $10 per million input / output tokens through Aug. 31, then $3 / $15.
Early testers praised its follow-through: bug fixes, pull requests, Salesforce updates, insurance workflows, legal research, and data exploration.
Anthropic says it has lower rates of hallucination and sycophancy than Sonnet 4.6, with cyber safeguards on by default.

How to try it:

Open Claude; Free and Pro users should see Sonnet 5 as the default.
In Claude Code, select Sonnet 5 for coding workflows.
For developers, call claude-sonnet-5 through the Claude API.

Why this matters: Sonnet is the model most Claude normies actually touch. Opus is the fancy chef’s knife, while Sonnet is the one that lives in the drawer and actually gets used for Tuesday dinner. Congrats Chef, you have a better everyday knife! Yes, Chef!

The “this is huge” camp is posting real task wins. One non-coder said Sonnet 5 helped him build five web apps in 10 minutes. Another user praised its agentic follow-through after it investigated a bug, wrote a reproducing test, fixed the issue, and verified the result without extra hand-holding. GitHub’s early Copilot tests also leaned positive, especially for CLI-style coding tasks.

The skeptical camp is staring at the bill. Theo and others flagged that Sonnet 5 can cost more than Opus 4.8 on some benchmarks once you count actual token usage, even if the sticker price is lower. Rohan Paul highlighted the same effective-cost problem. David Shapiro complained it goes off-task and lectures too much.

Although, TBH, a model that writes the bug fix and then gives you a TED Talk about why bugs are bad is very Claude-coded.

Our take: Sonnet 5 is Anthropic trying to make agents boring in the useful way: cheaper, safer, and default. But Fable 5’s return makes the timing awkward. Power users are already asking whether Sonnet 5 is the practical daily driver, or just the model you use while waiting for the restricted frontier model to come back. Don’t mind me, I’m just gonna smash that Claude app refresh button until the Fab 5 is back online…

FROM OUR PARTNERS

The best voice models, now across all channels

Most CX platforms do not own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, cost, and another vendor to manage.

ElevenAgents is the opposite. They make the voice models the market builds on, and ElevenAgents puts full orchestration on top. Voice, transcription, text-based chat, and reasoning run in one vertically integrated pipeline, so responses come back in <400 milliseconds and sound human, not synthetic.

Plus, you keep full control. Plug in any LLM, integrate tools, webhooks, and MCP servers, and ground responses in your knowledge base. Get an agent live in minutes, then A/B test with Experiments, enforce Guardrails, and version every change.

The payoff: more human conversations, lower latency, and far less time stitching infrastructure together. You build on the models you already trust. Pricing is transparent and flat at $0.08 per minute.

Start building free

🎓 AI Skill of the Day: Turn One Repeated Task Into a Claude Skill

If you keep re-explaining the same workflow to Claude, you’re doing the boring part manually. Package that routine into a reusable Skill instead.

A Skill is basically a saved instruction set Claude can follow again later: your role, process, rules, output format, and “please stop doing this weird thing” preferences. As Mr. Buzzoni pointed out, the people getting more leverage from Claude are moving from one-off prompts to repeatable systems.

Try this today:

Pick one recurring task you do weekly.
Write the messy version of your process: steps, examples, output format, and things to avoid.
Paste it into Claude with this prompt:

You are an expert at turning everyday workflows into reusable Claude Skills.

Here is one recurring task I do:

[PASTE YOUR WORKFLOW DESCRIPTION HERE]

Turn this into a complete, reusable Claude Skill with:
- a clear Skill name
- the role Claude should play
- the exact step-by-step process to follow
- the output format
- quality standards
- hard rules and things to avoid

Write it so I can save it and reuse it later without re-explaining the task.

Save the output somewhere handy. Next time, start with: “Use my [Skill Name] Skill on this input.”

Have a specific skill you want to learn? Request it here.

The Neuron AI Explained: AI Still Sees Like a Toddler. Andrew Dai Wants to Fix It

Click the image above to watch on YouTube

AI can write code, pass exams, and summarize the internet. But visual reasoning is still one of the biggest missing links in AI development: models can label what is in an image, then fall apart when they need to reason through diagrams, tangled cords, floor plans, robots, product designs, or the physical world.

In this episode, Corey and Grant talk with Andrew Dai, co-founder and CEO of Elorian and a former Google Brain/DeepMind researcher, about why today’s models still struggle with visual understanding and what “visual chain of thought” could unlock. Watch the episode here.

🍪 Treats to Try.

Google launched Nano Banana 2 Lite for faster, cheaper Gemini image generation and made Gemini Omni Flash available for developer video generation and editing.
Gemini image generation now gives more U.S. users personalized AI image tools inside Gemini, free to try.
Qwen-AgentWorld helps developers train and test agents in simulated environments like web browsing, Android, terminal work, search, and software engineering, free/open-source.
Ornith-1.0 gives developers open-source coding models that can create both solutions and their own test harnesses, free/open-source.
Browserbase Agents lets developers ship a browser agent from one prompt and one API call, with browser automation infrastructure built in.
X shipped an MCP server that connects Grok, Cursor, and Claude straight into the X API, so your agent can search the full post archive, check trends, manage bookmarks, and draft Articles using your own account's permissions.

📰 Around the Horn.

More below

OpenAI introduced GeneBench-Pro, a benchmark for AI agents doing real computational biology and genomics research.
Anthropic launched Claude Science, a beta research workbench with code-traced artifacts, on-demand environments, and 60+ optional scientific database connectors.
A new report says even $180K tech salaries no longer feel like enough in San Francisco, where AI money is pushing rents, home prices, and general spreadsheet despair into a new tax bracket.
Amazon launched a $1B Forward Deployed Engineering org to embed AI engineers with customers and build production agentic systems faster.
Etched exited stealth with a $5B valuation, $800M raised, and $1B in signed contracts for AI inference chips.
Meituan open-sourced LongCat-2.0, a large AI model trained on domestic Chinese chips.
Google Research introduced TabFM, a zero-shot foundation model for tabular data that beats tuned supervised models without per-table training or manual feature engineering.
Thinking Machines Lab and Bridgewater showed that expert investor annotations can fine-tune a smaller model to beat frontier models on real financial judgment tasks at lower cost.
Ramp and Revelio Labs found high-intensity AI spenders had faster headcount growth, complicating the “AI kills jobs” story.

FROM OUR PARTNERS

See Why HubSpot Chose Mintlify for Docs

HubSpot switched to Mintlify and saw 3x faster builds with 50% fewer eng resources. Beautiful, AI-native documentation that scales with your product — no custom infrastructure required.

Simplify Your Docs Today

📖 Midweek Wisdom:

Reinventing Entropy: Rohan Pandey, an ML researcher of seven years who wrote on compression and scaling laws, said 3Blue1Brown’s new video “still gave me fresh intuition on entropy.” Watch it for a visual explanation of why compression and intelligence rhyme (speaking of 3Blue1Brown, he just went on Dwarkesh as well!).
The Twilight of the Chatbots: Ethan Mollick argues AI is moving from chat interfaces into long-running agents that can complete work with less supervision.
Hermes Full Course: A 45-minute walkthrough for setting up Hermes as a 24/7 AI chief of staff, including Telegram, voice replies, Google Workspace, and recurring routines.
Cooking with OpenAI’s Research Chief: Mark Chen talks through AGI, o1, evals, scaling laws, and why public benchmarks can become less useful once everyone starts optimizing for them.
Matt Pocock’s Agentic Engineering Workflow: Matt Pocock argues the real edge in AI coding is the harness: better prompts, better skills, cleaner codebases, scoped tasks, and review loops.
AIEWF Daily Dispatch: Latent Space recaps day 2 of AI Engineer World’s Fair, sharing key insights from OpenAI’s Alexander Embiricos, Peter Steinberger, and more on “loops”, software factories, forward-deployed agent engineers, Warp’s factory pitch, and the open-model track from Z.ai and MiniMax.
How to Survive AI as a Non-Believer: Alberto Romero says your opinion of AI matters less than whether your workplace expects you to become fluent in it.