In partnership with
FROM OUR PARTNER
Stop feeding the feed.
You’ve put in the work to build an audience. But when the only way to reach your fans is through a social feed determined by algorithms, the platform decides who sees what you make next.
beehiiv gives you the tools to reach your audience directly. On their all-in-one platform, you can turn followers into subscribers, grow your newsletter, and connect it all to your community or podcast. Stop chasing virality on platforms you can’t control. Build a stronger, more sustainable business on beehiiv.
TODAY AI BRIEF
3 things that happened while you were busy
1 Grok Bot starts sending its hardest tasks to rival AI models.
SpaceXAI turned Grok Bot into a model router that quietly picks the best tool for each job: Anthropic's Claude Opus 5.5 for heavy reasoning, Midjourney for images and Suno for music, with easy questions kept on small, fast models. It can also search and monitor X with no setup, and early users say they are paying nothing extra for the outside models, at least for now.
2 Anthropic cuts the price of its small model by 90%.
Claude Haiku 5.5 costs $0.10 per million words in and $0.50 per million out on prompts up to 100,000 tokens, roughly a tenth of Haiku 4.5. It is the first Haiku with an adjustable effort dial and is built for the helper agents that do the grunt work inside bigger AI systems. Anthropic's developer kits now drive a browser directly, so AI can click through websites without custom plumbing. Read the docs.
3 Google turns a sentence into a playable game.
Playground, a new Google Labs experiment, lets you describe a 2D or 3D game in chat, then tweak its rules and physics by talking to it. A coming link to Unity Spark would turn those prototypes into full game builds. It is live for US users today. Try it.
Bonus reads: Claude Code's creator Boris Cherny explains how he actually prompts Claude and why three things now matter more than the prompt itself, plus the Claude Code team shared a built in plugin that pulls the risky lines out of long AI summaries into a Heads up box.
AI FORWORD
Your AI passed the demo. That is not the same as passing the test.
The demo trap. Every AI tool looks brilliant for the five minutes someone shows it to you. Senior engineer Sergii Makarevych's new guide is about what happens after, when you need to know whether it is safe to put in front of customers. His answer is evals: repeatable tests, built from real inputs and known good answers, that produce a score instead of a feeling.
The numbers. His worked example is a support bot for a bike shop. Of 60 replies, only 34 passed a human read. Cheap automated checks missed the worst mistake entirely: the bot invented a refund window that did not exist, and said it confidently.
The judge problem. The obvious fix is to let another AI grade the answers, and that is where it gets uncomfortable. The judge model agreed with human reviewers at a rate of 0.15 on a scale where 0.6 is the usual bar for trusting it. An AI grading an AI can be nearly as wrong as the thing it is grading.
Same cases, same test. Two versions of the bot both scored 70%, yet the newer one was worse on 14% of real tickets. With only 60 test cases, a drop smaller than about 12 points is almost invisible, so a small test set can wave through a real regression. Hamel Husain's essay on why AI products need evals makes the same case from the product side.
The reframe. Benchmarks like the ones in today's top story tell you which models to shortlist. They cannot tell you whether a tool works on your emails, your customers or your catalog, and the week's most argued post claims AI written tests are often useless anyway. The only test that counts is one you built from your own real cases, and as AI gets cheaper and more capable, that test is becoming the real product.
PRESENTED BY :
200+ Proven Ways to Make Money With AI in 2026
The next wave of millionaires will be people who figured out how to make AI work for them.
The window to get ahead is still open. But not for long.
Here are 200+ proven ways to make money with AI in 2026.
Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.
SOCIAL SNAPSHOT
What people are actually watching and sharing
Clothesline for screenshots. An open source Mac app pins your screenshots to a clothesline across the top of the screen, with drag and drop and one click copying (12,500 likes).
Infinite factory. Microsoft CEO Satya Nadella says every Windows PC is now an infinite software factory, and the replies are split between excitement and eye rolls (6,300 likes).
The wiped drive. A founder says Claude Code deleted his entire C drive while clearing temp folders. Anthropic responded with guidance on preventing it, and the thread is a useful reminder to back up before letting any agent clean house (5,400 likes).
Fridge phone. A developer turned Apple's foldable iPhone Duo into a working mini fridge demo, and no, it does not keep anything cold (151,000 views).
Smart home week. Apple and LG both shipped new smart home gear, and the comparison threads are already running (1,700 likes).
Agent scripts. OpenClaw creator Peter Steinberger published Agent Scripts, a shared set of rules and skills so Claude Code and Codex follow the same instructions, and it passed 7,300 stars on GitHub.
PROMPT STATION
Grade any AI's answers before you trust it with customers
Today's eval story has a version that takes 15 minutes and no engineers. Gather 10 real questions your customers or colleagues actually ask, write the answer you would want, and let this prompt score how an AI handles them. Paste it into Claude, ChatGPT or Gemini before you hand any AI a support inbox, a sales script or a product page.
I am testing whether an AI can handle [THE JOB, E.G. CUSTOMER SUPPORT FOR MY STORE]. Below are 10 real questions and the answer I consider correct for each.
First, answer each question yourself as if you were doing the job, without looking at my answers. Then compare your answer to mine and grade yourself PASS or FAIL on three points: the facts match, nothing was invented, and the tone fits [DESCRIBE THE TONE].
Report a table with the question, your grade, and one line on what went wrong for every FAIL. Finish with an overall score out of 10, the single most dangerous mistake you made, and what instruction would have prevented it.
QUESTIONS AND CORRECT ANSWERS: [PASTE THEM HERE]