Presented by
Have you ever checked what "powered by Claude" means on an app you use?
You order paneer tikka and get tofu. It looks the same and tastes close enough, and you still pay for paneer.
That's roughly where we are with AI. Plenty of apps say they run on Claude or GPT, and a few sites will sell you "premium AI" for a fraction of the official price.
And it has already happened.
Last year Moonshot, the company behind the open model Kimi, tested other companies selling its model, and some of them botched about 28% of tool calls (when the AI uses something like search or a calculator), while the original got none wrong.
With closed models like Claude or GPT, a middleman can't water down the model itself, but it can quietly send your question to a cheaper one.
Anthropic's own research names that exact risk: a model "secretly substituted in by a routing service for cost reasons."
Last week Anthropic showed a way to catch it. (writing this before diwali so..)
Below is the receipt it built, plus a 5-minute tofu test you can run on the AI tool you pay for today.
If you read till here, you might find this interesting
Presented by Angelino’s
I Almost Threw Out My Keurig. It Wasn't Broken.
Every cup from my Keurig started tasting weak, flat, stale — so I unplugged it and stuck it in a closet. Then I learned the dirty secret of grocery store K-Cups: by the time a pod reaches your kitchen, it's usually spent months going from factory to warehouse to truck to shelf. The flavor oils are long gone. Some brands even over-roast to hide it.
The fix wasn't a new machine. It was fresh pods from Angelino's, a family-owned L.A. roastery that roasts, packs, and ships within 3–5 days — straight to your door, from 39¢ a cup. Over 1,000,000 happy customers. 50+ flavors, no subscription required, and 15% off your first order, applied automatically at checkout.
Presented by Wispr Flow
Real-time dictation, built in-house, benchmarked against the big labs
Real-time speech recognition is harder than batch transcription. You don't get to wait for the end of the sentence. Wispr Flow built its own real-time model, Canto.
On a challenge set of messy audio (music, traffic, nearby voices, whispering, far-field speech), it had the lowest word error rate among the real-time models tested. It also tied for the lowest error rate on LibriSpeech. On 10 hours of real dictations from 2,300+ speakers, it beat every model tested, including Google, OpenAI, AssemblyAI and Deepgram.
Talk to Cursor, Claude Code or Slack, and get clean, formatted text as you go. Mac, Windows, iPhone and Android.
An AI answer, with a receipt
It's called FlashProve, and it attaches a receipt to an AI answer that proves which model wrote it, without the company showing anyone the model's insides.
Anthropic built it with cryptographers from EPFL, MIT and StarkWare, including one of the people behind Zcash, and put the code on GitHub for anyone to use.
Why it's not in your app yet
For now it's a research prototype, tested on Meta's open Llama 4 Maverick model, that handles chats of around 750 words and costs 70 to 95 times the computing it took to write the answer.
Anthropic has marked the code "not maintained," and it only tells you who answered, so a verified model can still make things up.
So until receipts show up, you're the receipt.
Run the tofu test
Easy questions won't help here, because even cheap models get them right. These three are built to trip up smaller models.
Run each one on the official app (claude.ai, ChatGPT or Gemini) and on the tool you pay for, twice each, since answers vary.
It has all three prompts, the answer key and tick-boxes, and it tells you whether your tool kept up.
Before you email support, check AI Nerfs.
It's a free tracker where people hit a button when Claude, ChatGPT or Gemini feels dumber than usual, and each model gets a score comparing the last 24 hours with its own normal.
When I checked, Gemini was running at 147%, so if your tool slipped on the same day, you're probably not the only one.
If the official app passes and your tool keeps slipping, email the tool's support team.
Ask which model and version answers you, and whether they ever switch to a cheaper one without telling you.
Don't ask the chatbot itself, since models often don't know their own version.
One honest caveat: this catches big downgrades, like a flagship swapped for a small model. It won't catch a quiet switch between two close versions, and that's the gap receipts would close.
Run it again every month or so, because these swaps rarely come with an announcement.
Hit reply with your results, and I'll share the most interesting ones next week.
Until next time,
Vaibhav 🤝
