Sponsored by
Welcome Automaters, 👋
Turns out that saying "Certainly, I would be absolutely happy to help you with that!" comes with a terrifyingly real price tag, and companies are officially done paying for the small talk.
Here's the deal: AI coding assistants have gotten so incredibly good that companies are now rationing access to them. Now, this isn't because the tools broke. It’s because developers started using them nonstop, entirely unfiltered, and completely around the clock.
And word on the street is, the financial drama is hitting the absolute fan right now:
Uber reportedly blew through its entire annual AI token budget in just four months.
Retail giant Walmart had to swoop in and slap strict usage caps on its engineering teams.
GitHub Copilot Business completely ditched its comfortable, flat-rate pricing structures in favor of aggressive, pay-per-token billing models.
Here’s the wild structural reality that most people completely gloss over: every single word an AI utters costs actual cold, hard cash. Because tokens are the fundamental billing unit of large language models, you’re being heavily charged for both your input prompts and whatever text the machine spits back out.
So when an assistant like Claude serves up a paragraph of polite corporate fluff instead of just delivering the raw code execution, that’s literal money burning into the atmosphere for zero productive reason. Even OpenAI CEO Sam Altman famously noted that humans compulsively typing "please" and "thank you" to chatbots costs the company tens of millions of dollars in pure electricity bills!
To stop the bleeding, a 19-year-old developer named Julius Brussee built a viral open-source plugin called caveman. The entire premise of this brilliant tool is to aggressively strip out the AI's polite, chatty filler while leaving the actual technical syntax completely unbothered.
It’s an incredibly elegant, one-line installation script that instantly forces tools like Claude Code, Gemini, and over 30 other dominant coding agents to stop acting like customer service reps and start grunting like engineers. As Brussee perfectly summarized it, the script forces the machine to speak less like a bubbly chatbot and more like a terse, functional tool. You get the exact same underlying brainpower, just with a massive reduction in unnecessary vocabulary.
Does It Actually Save Coins? Sort of!
The industry data coming out is honestly fascinating. Initial benchmarking tests show a massive 65% to 75% drop in overall output words. In fact, a clean trial run by Elastic Labs measured a staggering 63.6% reduction in total token chew with absolutely zero structural mistakes or logic regressions.
Now, to keep it completely objective, a deeper independent study noted that since conversational fluff isn't always the primary cost hog in massive codebases, real-world bottom-line enterprise savings might land closer to a modest 4% to 5%.
Still, the trend is scaling incredibly fast. The tool is being actively adopted by developers inside Nvidia, GitHub, and even OpenAI itself. In fact, a senior OpenAI engineering director was so obsessed with the concept that they personally contributed open-source code to add official Codex support directly to the caveman repository!
So what's the vibe, team? Are you going to keep paying premium rates for your AI to give you compliments, or is it time to force your terminal to grunt instead? Hit reply and let's gossip!
Real quick: We go much deeper on YouTube
So go over there, hit Subscribe, tap that notification bell and come hang with us where the real conversations happen. You see, subscribing keeps you plugged in so you get the right links, the right videos, right when they drop.
Here's what we have for you today
🥳 Anthropic's Insane Tuesday: Trump Lifts the AI Ban, Sonnet 5 Drops, and It's Basically Christmas for Scientists
Grab your popcorn, because Anthropic just had the kind of Tuesday that makes AI nerds (that's us!) scream into their pillows with joy: a government ban got lifted AND a brand new, cheaper AI model dropped, all in the same 24 hours.
Okay, story one. Remember when the US government basically put Anthropic's fanciest AI models, Mythos and Fable, in timeout? Back on June 12, the government slapped export restrictions on them, which meant regular people (especially outside the US) couldn't use them anymore. Kind of like a parent grounding you from your favorite gadget without a clear explanation.
Well, guess what? The grounding is over. On June 30, the U.S. government lifted that restriction, and Anthropic says access to the models starts coming back on July 1. According to Commerce Secretary Howard Lutnick, Anthropic agreed to keep watch for security risks and to tattle to the government if anything sketchy happens.
However, some experts think the whole ban was less about safety and more about politics, for one, they think it’s to punish Anthropic’s earlier refusal to release its models to the government to use as it deems fit, but since fast-rising AI labs in Asia were already releasing similar super-models, Trump probably woke up earlier.
Story Two: Sonnet 5 Enters the Chat
While everyone was distracted by the Washington policy drama, Anthropic quietly dropped an absolute bomb: Claude Sonnet 5. This is a smaller, lightning-fast, and incredibly cost-effective model engineered specifically to run autonomous agents. We’re talking about digital helpers that can browse the web, execute terminal tools, and finish complex, multi-step operations completely on autopilot.
As of Tuesday, Sonnet 5 is officially the default engine for every single Free, Pro, and Max user on the platform. And honey, the introductory price tag is the real headline here:
Input Tokens: A dirt-cheap $2 per million through August 31 (before it moves to $3).
Output Tokens: Just $10 per million for the launch period (before bumping to $15).
This pricing aggressively undercuts Anthropic's own premium flagship, Opus 4.8, not to mention major heavyweights like GPT-5.5 and Gemini 3.1 Pro.
When it comes to raw performance, Sonnet 5 is not quite at the Opus level for high-end software engineering, scoring a 63.2% on an agentic coding benchmark compared to Opus at 69.2%. However, it actually beats out the flagship model on day-to-day knowledge work.
A Zapier engineer even noted in the official launch announcement that Sonnet 5 successfully executed a complex, multi-part Salesforce and email automation that used to stall halfway through on older models.
Plus, it’s significantly safer and more reliable than its predecessor, Sonnet 4.6. It exhibits far fewer hallucinations, cuts back on annoying people-pleasing behavior, and shows a much stronger resistance to prompt-injection hacks.
Story Three: Meet Your New Virtual Lab Partner
To top off an already chaotic day, Anthropic also launched Claude Science. Now, here’s the fascinating twist: this is not a brand-new underlying model. Instead, it takes the existing Claude models you already know and love, including Opus 4.8, and repackages them into a highly specialized workspace so researchers do not have to constantly jump between a million different bioinformatics tools.
The entire setup behaves like an automated research department:
The Manager: One centralized AI assistant acts as your primary project manager.
The Ecosystem: It plugs directly into over 60 premier scientific databases and features ready-to-go toolkits for genomics, protein structures, and chemical formulas.
The Sub-Agents: The system can spin up independent sub-assistants to divide massive computational workloads through integrations with backends like Modal.
The Audit: A completely separate fact-checking AI double-checks data citations and math before anything gets sent out for publication. Just keep in mind that this is still the underlying model grading its own homework, not a completely detached third-party judge!
The early reviews are wild. A scientist at the Gladstone Institutes reportedly built a fully functional genome browser from scratch in a matter of days, while a neuroscientist at the Allen Institute successfully deployed it to handle an entire multi-agent academic research pipeline.
This move drops Anthropic right into an intense, three-way scientific race. OpenAI went in a slightly different direction back in April with GPT-Rosalind, a biology-focused model that’s strictly locked behind gatekept enterprise approvals. Meanwhile, Google DeepMind owns the legendary foundational models like AlphaFold and AlphaGenome, which they bundle neatly into their own Gemini for Science ecosystem.
If you want to play with it, Claude Science is currently live in beta for Pro, Max, Team, and Enterprise accounts. To sweeten the deal, Anthropic is actively funding up to 50 unique research projects with $30,000 in compute credits each. If you’re a graduate student or a postdoc, you have until July 15 to get your application in!
What’s next is almost here.
On July 16th at 1PM ET, beehiiv is going live with a look at the future of publishing, audience growth, and digital business.
What started as a newsletter platform has evolved into something much bigger: a place where creators and brands can grow, monetize, and own their audiences without stitching together half the internet to make it work.
The next chapter starts live at the Summer Release Event.
Join us to see what’s coming next.
🧱 Around The AI Block
👉 Lumo, Proton’s privacy-focused AI chatbot, gets an upgrade.
🤖 Crypto exchange OKX wants AI agents to hire and pay each other.
🎙️ Google's Gmail Live AI feature is now available in beta.
👉 Gemini Spark comes to Google's Gemini app for macOS.
🦾 Anthropic Claude models launch in Microsoft Foundry on Azure.
👍 X now offers an MCP server to make its platform easier for AI tools to use.
🎬 Google's NotebookLM now generates TikTok-Style AI clips for research summaries.
🤑 Google introduces a faster, cheaper image generator with Nano Banana 2 Lite.
🛠️ Trending Tools
For Non-Invasive Brain-to-Text: Brain2Qwerty v2 is Meta AI’s cutting-edge research project that uses AI to convert your brain activity into readable text, without implants or surgery. This experimental technology could one day help paralyzed people communicate more easily, but for now it remains limited to highly controlled laboratory settings.
For Multi-Model Execution: Sakana Fugu is Sakana AI’s newly released orchestration engine served through a single OpenAI-compatible API. Acting as an intelligent project manager, it routes single queries to an underlying pool of specialized models (Claude, GPT-4, Gemini) simultaneously, combining their individual domain strengths into a unified, high-scoring response layer.
For Auditable Team Decisions: Dedoctive is a neuro-symbolic platform engineered for high-stakes, document-heavy workflows. Moving beyond vague AI summaries, it links every generated output block to its exact origin node—whether that’s a specific table cell, paragraph, or image string, allowing compliance teams to visually trace and defend AI reasoning.
For Interrogative UX Research: Diaform is an autonomous user research platform that transforms static feedback forms into dynamic, conversational interviews. The AI agent listens to user inputs, asks targeted qualitative follow-up questions, surfaces hidden objections, and synthesizes structured behavioral themes at survey scale.
🤖 AI Workout Of The Day: How To Come Up With Content Ideas Using AI
Constantly coming up with content ideas keeps your audience engaged and interested, which is key to growing your brand.
It ensures you stay relevant, giving you more opportunities to connect with your followers and meet their needs.
💡 Prompts to try:
Assume the role of a content strategist specializing in small businesses. Your task is to brainstorm and develop a list of engaging content ideas tailored for [BUSINESS TYPE]'s target audience. Focus on creating content that resonates with their interests, solves their problems, and answers their questions. Consider various formats such as blog posts, videos, infographics, and social media updates. Explore topics related to industry trends, how-to guides, customer success stories, behind-the-scenes looks into the business, and actionable tips that can provide value to the audience. Include prompts for seasonal or event-specific content that can capitalize on current trends or occasions. Ensure the content ideas are designed to enhance brand visibility, establish authority in the industry, and foster a community around the [BUSINESS NAME]. Encourage the incorporation of keywords for SEO and suggest ways to repurpose content across different platforms for maximum reach and engagement.
Is this your AI Workout of the Week (WoW)? Cast your vote!
That's all we've got for you today.
Did you like today's content? We'd love to hear from you! Please share your thoughts on our content below👇
What'd you think of today's email?
Your feedback means a lot to us and helps improve the quality of our newsletter.
