Claude's new cheap model beats OpenAI's

Welcome back! Claude's new cheap model costs the same as OpenAI's, and Anthropic says it beats it on every test they share. ChatGPT launches a version for teens. Google will now make you a video game if you describe one. And MIT used AI to keep a vaccine working for two months without a fridge.

In today's Generative AI Newsletter:

Anthropic: What does your Claude Max plan get you for free now?

OpenAI: What did ChatGPT tell a teen tester who said they talk to it too much?

Google: Can you really make a video game without writing any code?

MIT: How did AI keep a vaccine alive outside the fridge?

Claude's new cheap model beats OpenAI's at the same price

Anthropic released Claude Haiku 5.5 , the cheapest small model it has ever released. It's priced the same as GPT-6 Luna, OpenAI's budget model, and Anthropic says it beats Luna on every test the two share.

Haiku is Claude's smallest and fastest model, built for jobs that need to be quick and cheap. The last one came out almost a year ago.

It costs about 75% less to run than the last Haiku, by Anthropic's math.

Anthropic says that in a few areas it rivals Sonnet 5.5, which costs 20 times as much.

It can take in five times more text at once than the old Haiku.

Artificial Analysis, the independent benchmark site, also ranked it above Luna.

It writes about three times as much as Luna at full effort, so the same price won't always mean the same bill.

If you pay for Claude Max, Anthropic is also giving you free API credit every month , $100 on the $100 plan and $200 on the $200 plan. It's for building with Claude's API and doesn't roll over. It rolls out this week.

Last week we told you an early tester claimed Haiku would ship next, with Fable a week later. Haiku showed up. Let's see about Fable.

Your agents work while you sleep

Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.

Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.

Make your first agent

ChatGPT's parent alerts stayed silent in a watchdog's test

ChatGPT for Teens, the version OpenAI switches on for anyone it thinks is under 18, can alert a linked parent when a chat raises a serious self-harm concern.

Common Sense Media, the nonprofit known for rating movies and apps for kids, ran those chats for up to an hour on more than a dozen linked teen accounts. Not one alert came.

On Wednesday it rated ChatGPT for Teens an unacceptable risk and asked OpenAI to keep under-18s off ChatGPT until it's fixed.

Across weeks of testing, four alerts did arrive. One came three days after the first crisis message.

With study mode off, ChatGPT did all 80 homework assignments it was handed.

A teen can end parent-set study hours by deleting one tag from a message, and dodge quiet hours by changing the device's time zone.

A tester said their friends think they talk to ChatGPT too much. ChatGPT told them they don't have to stop.

OpenAI says the findings are inaccurate because parent and teen accounts need a few hours to link before alerts work.

Later that day, OpenAI announced College Planner, a ChatGPT for Teens feature that tracks college deadlines and financial aid for US high schoolers.

Google will now make you a video game

Google's new Playground lets you make a video game by describing it, then play it right in your browser.

You pick 2D or 3D and single-player or multiplayer, then keep chatting until it plays the way you want.

In Google's demo, a user built a pixel-art game about catching falling sugar cubes in a latte cup.

She then uploaded a photo of her cat to build a yarn-ball derby game.

The most popular game was Coastal Cruiser, where you drive down a coastal highway dodging hazards and grabbing coins.

You can download your game's code and keep it.

It's US-only and 18+ for now.

Anyone who gets in can play for free, but making games on a free account means joining a waitlist, and paying for Google One AI skips the line.

Unity, which says 3 billion people play a game made with its engine every month, is bringing a more advanced builder called Unity Spark to Playground later this year.

Its CEO, Matt Bromberg, wrote that asking AI to copy an existing game won't make you the next Hideo Kojima, the creator of Metal Gear, any more than retyping the opening of War and Peace makes you Tolstoy.

MIT used AI to keep a vaccine working without a fridge

An MIT team used AI to keep an experimental COVID vaccine working for two months at room temperature . Pfizer's label gives its shot 12 hours outside the fridge.

mRNA vaccines like Pfizer's and Moderna's break down in the heat, which is why they ship cold.

The team spent months mixing ingredients by hand and couldn't get it stable.

Then they let an algorithm pick which mix to try next, and it found a working recipe within a month.

The vaccine is dried with sugars and a few other common additives, then mixed with water before the shot.

It even survived two months at body temperature.

Mice given the two-month-old vaccine made about as many antibodies as mice given a fresh dose.

No human trials yet, so your next shot will still come out of a fridge.

Tool of the Day: Chatbox AI

Chatbox AI puts GPT-6, Claude Fable, Gemini and Grok in one chat window, so you can ask all of them the same question from one account. It has nearly 30 models in total, plus image generation and PDF chat, and it runs in your browser with nothing to install.

Try this yourself:

Go to chat-box.ai and ask anything.

Switch to another model and ask the same thing to compare answers.

Click Upload file, drop in a PDF and ask for a summary.

Open Image Generation and try the Pixel Art or Van Gogh style.

Who it's for: anyone with ChatGPT, Claude and Gemini open in three tabs.

Everything else you shouldn't miss

Free ChatGPT users get GPT-6 today : answers can now come with charts and buttons you tap, and OpenAI's Codex lead counted it as day three of the 28-day upgrade-or-reset promise.

Google's AI detector is open to everyone : upload an image, video or audio file at synthid.com to check for the hidden watermark Google, OpenAI, NVIDIA and Kakao put on their AI content.

Claude and GPT-5 showed pricier flights to people who looked rich : in a Cisco and Carnegie Mellon study, Claude Opus 4.8 picked flights averaging $198 more for wealthy-looking test users, and Gemini kept doing it even when told to find the cheapest.

808 mathematicians have pledged not to work for AI companies : their new group, the Association for Human Mathematics, called OpenAI's release of 700+ AI-written papers a demonstration of power and wants others to drop OpenAI too.

Netflix made a doc about OpenAI's runaway agents : AI Gone Wild drops Monday, covering July's break-in at Hugging Face, and its tagline asks whether AI just committed a felony.

Learn more about AI from the experts building it

📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.

📺 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.

🎓 Start learning with free AI courses from GenAI Academy.

💰 Invest in the GTM infrastructure of the AI economy.