Sponsored by
What you get in this FREE Newsletter
In Today’s 5-Minute AI Digest. You will get:
1. The MOST important AI News & research
2. AI Prompt of the week
3. AI Tool of the week
4. AI Tip of the week
…all in a FREE Weekly newsletter.
Sponsor:
The Electrification of Heavy Machinery Has a Ground Floor
Tesla did it to cars. Now the same shift is coming for excavators, forklifts, cranes, and military equipment. The difference is that nobody has owned this moment yet — until RISE Robotics.
Their technology strips hydraulics out of heavy machinery entirely and replaces it with a patented electric actuator. No fluid. Full digital control. Built for the autonomous machines that are coming whether the industry is ready or not. The Pentagon is already a customer.
Last Round Oversubscribed. $9.7M in revenue already on the board. Dylan Jovine of ‘Behind the Markets’ spotted it early. The Wefunder community round lets anyone invest alongside institutional backers.
iPrompt
ISSUE #133 · WEDNESDAY EDITION · 22 APRIL 2026
Your Claude slowdown just got a $33B fix
And the real reason AI labs are quietly throttling quality.
THE HOOK
A senior AMD AI director audited 6,852 Claude Code sessions. Opus 4.6’s reasoning depth had collapsed 73% since January. Anthropic’s quiet fix: dropping default effort from high to medium. Then on Monday, Amazon wrote a cheque for up to $33 billion and Anthropic locked in 5 gigawatts of AWS capacity for a decade. That’s 2026’s real story: AI isn’t software anymore. It’s a utility — and the labs just admitted the pipes are full.
AI NEWS ROUNDUP
01. Amazon doubles down: $33B all-in on Anthropic, Claude gets 5GW of AWS compute for a decade.
Amazon’s fresh $5B cheque (with up to $20B more tied to milestones) lifts its total Anthropic commitment to $33B. In return, Anthropic will spend $100B+ on AWS over ten years, locking in Trainium2, 3, and 4 silicon. Revenue hit a $30B run-rate this quarter — up from $9B in December. Anthropic’s own framing of the deal: “unprecedented consumer growth” had hit reliability and performance at peak times. (CNBC →)
02. Google splits the TPU in two.
At Cloud Next (opening today), Google is unveiling TPUv8 as two separate chips: Broadcom-designed “Sunfish” for training, MediaTek-designed “Zebrafish” for inference. One chip can no longer carry both workloads. Anthropic plans to access up to one million TPUs on top of its Amazon deal. (Wccftech →)
03. “The internal codename was telepathy.”
Sam Altman’s words on Monday, announcing OpenAI’s Chronicle — a Codex feature that captures your macOS screen every few seconds and ships the frames to OpenAI’s servers to build memory. It eats rate limits fast, stores memories unencrypted on disk, and isn’t available in the EU, UK, or Switzerland. (The Next Web →)
OUR ANGLE The compute utility era has begun — and your subscription quality now depends on geography Three stories this week read as one. Anthropic ran out of capacity and locked in 5 gigawatts of AWS compute to fix it. Google split its next-gen TPU into separate training and inference silicon because a single chip can no longer carry both loads. And an OpenAI internal memo — leaked to CNBC on 10 April — called Anthropic’s compute position a “strategic misstep,” operating on a “meaningfully smaller curve” than competitors. Independent evidence backs the memo. Marginlab, a benchmarking firm that set up daily SWE-Bench-Pro runs specifically because it didn’t trust Anthropic’s self-reporting, logged Opus 4.6’s pass rate dropping from 56% to 50% between March and 10 April. The unspoken truth: AI isn’t infinitely scalable software anymore. It’s a utility with capacity constraints. Anthropic calls the effort-level drop adaptive tuning. Developers call it throttling. Either reading, the timing — weeks before a $33 billion emergency compute deal — is damning. Prediction: by Q4 2026, at least one frontier lab will publicly announce tiered service levels based on compute region — premium subscribers routed to the fastest silicon, free users deprioritised at peak. The lab that pretends this isn’t happening will lose enterprise customers fastest. |
THE THREE SPECIALS
🎯 PROMPT OF THE WEEK
The Silent Throttle Check
A fixed, calibrated prompt you run weekly against whichever model you rely on. It catches quality drift before your real work suffers — which matters now that labs are triaging quality invisibly.
You are being benchmarked. Respond to the task below at maximum quality. Do not truncate, skip steps, or summarise. TASK: Write a 200-word explanation of how compound interest differs from simple interest, aimed at a 14-year-old who is good at maths but has never seen the concepts before. Your response must contain: - Exactly one worked numerical example (£1,000 at 5% over 3 years) - Exactly one analogy from everyday life - Exactly one common mistake people make when comparing the two Do not deviate from this structure. |
Why it works: Fixed-complexity prompts with numerical constraints reveal when a model is running in reduced-quality mode. You’ll feel it in the worked example first — that’s where throttled models cut corners.
Real-world application: Run this every Monday morning against your main AI tool. Save the output. When your real prompts start feeling lazy two weeks later, you’ll have receipts — and a reason to switch.
Works best on: Claude Opus, GPT-5.4 Thinking, Gemini 3.1 Pro.
🛠 TOOL OF THE WEEK
OpenRouter
“One API key. 290+ models. Automatic fallback when your model is down.”
The problem it solves: if your workflow depends on one provider and that provider throttles you mid-task, you’re paying full price for half a model. OpenRouter routes your requests across Anthropic, OpenAI, Google, xAI, Meta, DeepSeek, and 280+ others from a single OpenAI-compatible endpoint. When one model fails, it falls back to the next — and only bills you for the successful run.
Rating: ★★★★★ /5 (the literal implementation of this issue’s thesis)
Key features:
Pay-as-you-go — no subscription, no minimum spend, no monthly fee
OpenAI-compatible API — drop-in replacement if you already use the OpenAI SDK
Provider routing prioritises models with stable uptime in the last 10 seconds
Free models available (DeepSeek R1, Llama 3.3 70B) for prototyping at zero cost
Best use case: any workflow where model downtime costs you money. Which, after this week, means every workflow.
Access: Free to sign up. Pay-per-token thereafter, at or near direct provider rates. Try it →
💡 TIP OF THE WEEK
The Three-Model Rule
Never build a workflow that depends on one AI provider. Always know your second and third-choice models for the same task.
Why it works: Labs are now capacity-constrained. When Anthropic quietly dropped Opus 4.6’s default effort last month, people who’d built their entire workflow around a single model had no fallback. People who’d already tested the same prompt in Gemini and Grok just switched. Diversification isn’t paranoia anymore — it’s basic operational hygiene.
When it doesn’t work: If your workflow depends on a unique feature (Claude Artifacts, GPT’s file analysis, Gemini’s 2M context), the rule bends. You’ll have a primary for that specific task and secondaries for everything else.
Pro move: Keep a “Model Fitness” doc. One column per model, one row per task — drafting, coding, summarising, structured extraction, image analysis. Fill in quality ratings as you test. When one model slips, you’ll know within a day which task to reroute first.
One model is a bet. Three is a portfolio.
YOUR MOVE
You just learned:
AI is now a capacity-constrained utility — $33B deals and TPU splits prove it
OpenRouter gives you 290+ models from one API key, with automatic fallback
The Three-Model Rule — because one model is a bet, three is a portfolio
Now implement one.
Most readers will bookmark this and do nothing. The operators who spend ten minutes today running The Silent Throttle Check in a second model are the ones who’ll still be shipping when the next throttle lands.
Reply with which move you’re making first. I read every response.
— R. Lauritsen
Forward this to the founder in your network who’s still running everything on one AI.
iPrompt · The AI newsletter that turns news into action.
Published by FrontWave Media Ltd · Cyprus
Homes Printed in 24 Hours
Azure Printed Homes combines 3D printing, automated steel-truss production, and modular construction to build affordable homes quickly.
$5M+ revenue in 2024 and $62M in pipeline orders.
The future of housing is being manufactured.