TL;DR
Anthropic cuts Haiku 5.5 prices up to 90% as OpenAI puts GPT-6 in front of 1.2B ChatGPT users
Figma's design agent goes generally available
Google launches Playground, a prompt-to-game platform
Mistral previews Large 4, a 1T-parameter open model
Claude plus Grok Bot can staff your business with AI employees
FounderRead turns what founders are reading into a feed
What does a new AI model actually cost?
When a new model launches, most people repeat the press release. Karo Zieminski of Product with Attitude does something different: she builds with it and shows you the receipts.
Her latest piece looks at Jev, a new kind of model that doesn't chat. It makes small, structured decisions, like sorting an email or scoring a headline. Karo ran it on her own account, measured what it actually cost per request, and built a working SEO scoring dashboard for her newsletter in about four minutes.
Karo tries the tools, builds real stuff, and tells you what worked. Her newsletter, Product with Attitude, is a good one to have in your inbox.
ANTHROPIC JUST MADE ITS SMALLEST MODEL 90% CHEAPER
Anthropic released Haiku 5.5 on Wednesday, and input tokens now cost a dime per million. Yes, a dime. On prompts under 100K tokens, output is $0.50 per million, down from $1 and $5 on Haiku 4.5. Above 100K tokens, the cut is 50%, and Anthropic says the average bill drops about 75%.
That matches OpenAI's GPT-6 Luna on short prompts, and Luna just got a lot more important. OpenAI is rolling GPT-6 out to all of ChatGPT, with Luna covering Free and Go users across roughly 1.2 billion weekly users. It also shipped "Intelligent UI," where a reply can show up as buttons, forms, charts, or a mini tool instead of a wall of text. So the small-model price war is now a fight over what powers the default AI experience for a huge slice of the internet.
Anthropic says Haiku 5.5 beats Luna on most of its own benchmarks, including OSWorld and Terminal-Bench. It has a 1 million-token context window, up to 128K tokens of output, and an adjustable effort setting. It still trails Sonnet 5.5 on hard agentic coding, and Sonnet's cache-read pricing got cut in half too.
Small models do the grunt work in agents: routing, extraction, subtasks. A single task can fire hundreds of those calls, so a 75 to 90% price drop changes which products are worth building. Ideas that got shelved last month for cost reasons might pencil out now.
OUR TAKE
Frontier models get the headlines. Cheap models get the margins and the users.
If you're building on agents, rerun your cost math this week. Anything you shelved for being too expensive might work now.
If you're a thin wrapper, you're getting squeezed from both sides. Model costs are falling, and ChatGPT can now generate the UI on its own.
QUICK HITS
Figma's design agent just got a real job: After a beta run that started in May, the agent is now generally available and works on the canvas with your actual components, variables, and team files instead of flat images. It can generate layout directions, bulk-edit across screens, turn comments into changes, and build prototypes from a prompt. New this round, you can give it guidelines to follow and watch what your teammates' agents are doing.
Europe's big AI hope is named Le Chonk: Mistral just previewed Large 4, a 1-trillion-parameter open-weight model (49B active) that it calls the strongest open model built outside China by a wide margin. It's still behind the frontier in areas like coding, and only developers, cybersecurity leaders, and state authorities get it for now, with a wider release due later this month.
Google's new game engine is a text box: Google just launched Playground, which turns a typed description into a playable game (US-only for now, and how much you can create depends on your Google One tier). Google is already pointing "dedicated creators" to a Unity Spark integration, which hints the prompt-only version tops out somewhere.
28 Wishes sells ice cream in LA. Below 70°F, sales fall about 20%. So they put $20 a day into Kalshi weather markets, taking the cold side. The days that keep customers away now pay something back. Learn more.
HIRE A WHOLE TEAM WITHOUT POSTING A JOB
Use Claude to design your org chart, then have Grok Bot staff it with AI employees that already know your voice.
⏱️ 1-2 hours | 🔧 Needs: Claude, a Grok Bot account, Google Drive, your existing Claude skills (optional)
Why build this? Delegating takes weeks of job posts, hiring, and training. A lone chatbot forgets your voice every session. Here, Claude designs the team and Grok Bot staffs it with AI employees that share your files and sound like you.
Steps:
Have Claude design the team. Prompt: "Based on my business, write a prompt my AI COO can use to create managers and employees, with responsibilities, key tasks, and SOPs for each department." List your departments (sales, client acquisition, content, scaling).
Create a COO in Grok Bot. Make a new bot, name it, and paste Claude's prompt. It will ask a few questions about your business, then return a full org chart.
Edit the org chart before approving. Cut roles that shouldn't be AI (like a closer) and add what's missing (like a short-form script writer). Then tell the COO to build the team.
Connect Google Drive and clone your voice. Authenticate Drive so every employee can read your Docs and Sheets. Point your script writer at your best-performing scripts and tell it to learn how you talk and fix your bad habits.
Port your Claude skills. Upload your SKILL.md files and tell the COO to assign each one to the right employee. Script writing skill goes to the script writer, repurposing goes to the distribution editor, and so on.
Test with a real task and give feedback. Ask for an actual script, tell it what's weak (like the hook), then say "update your memory" so the fix sticks.
Expected outcome: A working AI team with defined roles, your voice, and your Claude skills built in. Start with content, review the output like you would a new hire, then add roles once the first ones are solid.
Full tutorial:
TOOL OF THE DAY
A book club for founders, built around X.
Founders love saying they read a lot. FounderRead makes it public. Connect your X account (it's free), log books as you start and finish them, rate them, and share updates to X in one click. The home feed shows what other founders are reading as it happens, there are curated shelves for things like finding product-market fit, fundraising, and hiring
