I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x and Subscriptions are Dead. | GMP EP03

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x and Subscriptions are Dead. | GMP EP03

AI token costs have crashed 850x since 2020 — what cost $60 per million tokens now costs 6 cents — and one host is burning $12,500 worth of tokens a month on a $200 subscription.

Mar 9, 2026 1:19:09 Difficulty: Intermediate Played

TL;DR

Three AI practitioners dissect the week's biggest stories: Anthropic's rejection of a Pentagon surveillance contract (and OpenAI's same-day swoop to claim it), the deep ties between effective altruism and Anthropic's decision-making, and a jaw-dropping chart showing AI token costs have fallen 850x since 2020. They also debate whether running AI locally on a Mac Studio makes sense (verdict: not yet), the collapse of AI avatar startup Icon.com versus Higgs Field's explosive growth, and GPT-5.4's surprising computer-use benchmark. Key takeaway: plan your AI system architecture before you build it, or you'll rebuild it from scratch.

#AI government contracts #Anthropic ethics #effective altruism #token cost deflation #open-source AI models #AI hardware local inference #Claude Code agentic AI #GPT-5.4 benchmarks #AI subscription economics #AI video generation #SaaS competition AI era #AI productivity debate #chip demand investing #AI avatar startups #Pentagon AI deal #Anthropic #OpenAI #Claude #government contract #Pentagon #token cost #AI pricing #open source AI #Claude Code #GPT-5.4 #AI hardware #Mac Studio #Higgs Field #SanDisk #Samsung #vibe coding #AI agents #subscription model #Gemini

Three AI practitioners debate Anthropic's Pentagon contract rejection, OpenAI's government deal, effective altruism's grip on Anthropic, the 850x crash in token costs, and the real-world tradeoffs of running AI locally versus in the cloud.

Chapter list
OAuth (Open Auth)
An open standard authentication protocol that lets third-party apps access a user's account without sharing passwords — here, used to tap a paid AI subscription outside its intended platform.
Pre-training
The initial, large-scale phase of training an AI model on massive datasets before any fine-tuning or specialisation; likened in the episode to going to school before an exam.
Post-training
Fine-tuning steps applied after pre-training — including RLHF and instruction-tuning — to align a model's behaviour; compared to coaching after school.
RLHF
Reinforcement Learning from Human Feedback: a technique for fine-tuning AI models by rewarding outputs rated positively by human evaluators.
Effective Altruism (EA)
A philosophy that uses evidence and reasoning to determine the most impactful ways to benefit others — central to Anthropic's founding culture and funding network.
Open Philanthropy
A major EA-aligned philanthropic fund co-founded by Dustin Moskovitz; one of Anthropic's earliest and largest backers.
Claude Constitution
Anthropic's published set of principles that govern Claude's behaviour — a strict guideline the model is not supposed to deviate from, closely tied to EA philosophy.
Claude Code (OpenCLO / OpenClaw)
Anthropic's agentic coding assistant that operates in terminals and IDEs; also used in the episode to refer to the broader open-source framework for routing AI tasks.
Vibe coding
A colloquial term for rapidly prototyping software by prompting AI models iteratively, relying on the AI to write most of the code.
Distillation
A technique where a smaller 'student' model is trained to mimic the outputs of a larger 'teacher' model, making open-source models competitive by learning from frontier closed-source ones.
Agentic AI
AI systems that can autonomously plan, take actions, and call tools over multiple steps to complete complex tasks without constant human prompting.
VPS (Virtual Private Server)
A virtualised cloud server that acts like a dedicated machine, giving users more control over their data than shared hosting or consumer AI apps.
ARR (Annual Recurring Revenue)
A standardised metric for subscription businesses representing the annualised value of recurring revenue — used here to gauge Higgs Field's growth trajectory.
ASIC
Application-Specific Integrated Circuit: a chip designed for one specific task; here used loosely to refer to specialised AI accelerators in the cloud.
TPU
Tensor Processing Unit: Google's custom AI chip, giving Google a full-stack hardware advantage over Anthropic and OpenAI who rely on third-party silicon.
OSWorld
A benchmark that tests an AI agent's ability to navigate and operate a real computer desktop — GPT-5.4 reportedly surpassed the human baseline on this test.
Moore's Law
The observation that the number of transistors on a chip doubles roughly every two years, roughly halving cost — invoked here to frame the 850x drop in AI token prices.
FUD
Fear, Uncertainty, and Doubt: negative sentiment or disinformation spread about a product or company, often to damage reputation or suppress adoption.
Loss leader
A product sold below cost to attract customers and gain market share, with the expectation of profiting later — applied here to AI subscription pricing.
Quant (QUANT model)
A quantised version of a large language model compressed to run on consumer hardware with reduced memory requirements, at some cost to output quality.

Chapter 3 · 03:05

Topic 1: Every Time Anthropic Draws a Line, OpenAI Picks Up the Contract

Anthropic missed a 5PM Pentagon deadline. By 11PM, Sam Altman had tweeted the deal.

Technology
Why Anthropic Walked Away From the Pentagon Deal

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

The Pentagon gave Anthropic a hard 5PM deadline to accept unrestricted military use of Claude. Dario was in meetings and missed it. By 11PM, Sam Altman had tweeted a done deal with OpenAI. What looked like a principled stance was also, at least partly, a scheduling failure — but the PR outcome for Anthropic was paradoxically positive.

Chapter 4 · 09:20

Effective Altruism, Anthropic's DNA, and the SBF Connection

Anthropic's CEO, COO, investors, and major donors are all one EA social network.

Society & Culture
Effective Altruism Is the Hidden Operating System of Anthropic

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Society & Culture

Anthropic's founders, funders, and key executives are almost all connected through the EA movement. Dario's college roommate co-founded Open Philanthropy, married Dario's sister Daniela, and joined Anthropic in 2025. SBF's $500M investment used FTX customer funds. EA isn't just Anthropic's philosophy — it's its entire social graph.

Chapter 5 · 21:10

SBF's $500M Bet: A 30x Return That Never Was

I pay $200/month for AI. I'm actually using $12,500 in tokens. The math doesn't work.

Technology
$200 Subscription, $12,500 in Actual Tokens — The Subscription Is Already Dead

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

Ben pays $200/month for Claude Max but burns roughly 500 million tokens a month — equivalent to $12,500 at API rates. AI providers priced subscriptions like a normal buffet, but developers are now taking entire restaurants' worth of food home. The OAuth crackdown from Anthropic and Google is the inevitable response.

Chapter 6 · 22:20

Subscription Arbitrage Is Dead: The $200 vs $12,500 Token Gap

Ben pays $200/month for a Claude Max subscription but consumes roughly 500 million tokens — equivalent to $12,500 at API pricing — demonstrating how broken subscription economics are.

Chapter 7 · 26:15

850x Token Cost Crash: From $60 to 6 Cents Per Million Tokens

AI tokens cost 850 times less than they did in 2020. This changes everything.

Chapter 8 · 28:50

Open Source vs Closed Models: Privacy, China, and Distillation

Every time OpenAI improves, the open-source models that compete with it get better too.

Chapter 9 · 32:30

Google Gemini's Pricing Edge and the Sleeping Giant Thesis

Claude Code already accounts for 4% of all GitHub commits and is projected to reach 20% by year-end — a remarkable rise for a single AI coding tool.

Chapter 10 · 34:50

Claude Code Hits 4% of GitHub Commits — Projected 20% by Year-End

Claude Code's framework just became the most-starred GitHub repo ever — bigger than React.

Technology
OpenClaw Is the Most-Starred GitHub Repo Ever — Bigger Than React

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

Claude Code's open-source framework has surpassed React as the most-starred GitHub repo of all time. Luca reads this as a signal that the winning consumer AI product won't come from a big lab — it'll be a closed, polished version of OpenClaw: model-agnostic, self-hosted, and shipped out of the box.

Chapter 11 · 36:30

OpenClaw: The Most-Starred GitHub Repo and the Future Consumer AI Product

What if the Jony Ive AI device just shipped with Claude Code pre-installed? That's the killer app.

Technology
The Jony Ive AI Device: What If It Shipped With OpenClaw Pre-Installed?

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

Jony Ive and Sam Altman are reportedly building an AI device together. Luca's provocative take: the killer device isn't some futuristic gadget — it's one that ships with a secure, pre-configured version of Claude Code's open-source framework, letting anyone run personalised AI agents out of the box without cloud dependency.

Chapter 13 · 42:30

Local AI Hardware Reality Check: Mac Mini vs Mac Studio vs Cloud

Mac Studio fans: 34 seconds for 400 words locally vs 7 seconds in the cloud. Cloud wins.

Technology
Local AI vs the Cloud: The Mac Studio Case Doesn't Add Up

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

Running a 32B quantised Llama model on a Mac Mini takes 34 seconds to write 400 words. Claude Sonnet in the cloud does it in 7 seconds. And the local model is meaningfully dumber. Unless you've fully systematised your workflow around dedicated hardware, local AI inference is slower, weaker, and barely more private.

Chapter 14 · 47:00

Icon.com Bankruptcy vs Higgs Field's Billion-Dollar Trajectory

One AI startup spent $12M on its domain and went bankrupt. The other is heading for $1B ARR.

Business
Icon.com Spent $12M on a Domain and Went Bankrupt; Higgs Field Is Heading for $1B ARR

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Business

Icon.com raised big money from Peter Thiel and OpenAI, spent $12M on a domain, launched a chaotic landing page, and is apparently bankrupt. Higgs Field took a disciplined approach to AI video generation, grew from 15M users to 4.5M daily videos, and is on track for $1B ARR. Same AI wave, completely different execution.

Chapter 15 · 53:10

Cluely's Revenue Scandal, Pivots, and Coming Back Down to Earth

Cluely's founder Roy publicly claimed $7 million ARR to TechCrunch when the company was actually doing closer to $4–5 million, issuing a public retraction describing it as 'the only blatantly dishonest thing I've said online.'

Chapter 16 · 57:30

GPT-5.4 Drops: Computer Use Benchmark Breaks the Human Baseline

GPT-5.4 just clicked its way past the human baseline on real computer use benchmarks.

Technology
GPT-5.4 Just Crossed the Human Baseline on Computer Use

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

GPT-5.4 jumped from 47% to 75% on OSWorld, a benchmark testing real desktop navigation, in a single release — crossing the human baseline. It apparently clicks around the internet at human speed. Previous agent modes looked impressive but were painfully slow; this model may be the first where computer use is genuinely practical.

Chapter 17 · 1:04:00

AI Bots, Web Blocking, Worldcoin, and the Identity Problem

SanDisk went 20x in a year. AI is a memory business — and most people missed it.

Chapter 18 · 1:06:50

SanDisk 20x, Samsung 250%: The Memory Stocks Riding the AI Wave

SanDisk's stock rose from a low of around $30 to $600 — a roughly 20x gain in a single year — driven by surging AI demand for memory storage.

Chapter 19 · 1:08:50

Is AI Actually Making You More Productive? Honest Answers from the Trenches

AI made you more productive. It made your competitors more productive too. Net result: harder.

Business
AI Productivity Is Real — But It Made Everything More Competitive, Not Easier

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Business

AI genuinely makes individuals more productive. But it has also turbo-charged every competitor. Landing pages are better, content is cheaper, and growth channels are flooded. The net result in SaaS: more capability, more competition, similar difficulty. The humans are still winning on outreach and relationship-building.

Chapter 20 · 1:16:10

Closing Advice: Plan Your AI System First, Then Build

Zoom out before you build with AI — or you'll spend two weeks building something you delete.

Technology
Plan Your AI System Before You Build It — Or You'll Rebuild From Scratch

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

The most common AI workflow mistake is skipping the architecture phase. Rick rebuilt his first vibe-coded project from scratch in 48 hours after two weeks of messy iteration because he hadn't planned the full system. Map every step, every agent connection, and every tool integration before writing a single prompt.

No indexed bits in this chapter.

Show stoppers

Technology
$200 Subscription, $12,500 in Actual Tokens — The Subscription Is Already Dead

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

Ben pays $200/month for Claude Max but burns roughly 500 million tokens a month — equivalent to $12,500 at API rates. AI providers priced subscriptions like a normal buffet, but developers are now taking entire restaurants' worth of food home. The OAuth crackdown from Anthropic and Google is the inevitable response.

Technology
Why Anthropic Walked Away From the Pentagon Deal

I Pay $200. I Use $12,500 in Tokens. AI Costs Crashed 850x … · Mar 9, 2026 Technology

The Pentagon gave Anthropic a hard 5PM deadline to accept unrestricted military use of Claude. Dario was in meetings and missed it. By 11PM, Sam Altman had tweeted a done deal with OpenAI. What looked like a principled stance was also, at least partly, a scheduling failure — but the PR outcome for Anthropic was paradoxically positive.

Snapshots ()

Key Quotes ()

This episode

Claims & Sources

2 / 20 cited (10%)

Factual claims made this episode, and whether a source was named.

The Pentagon's deadline for Anthropic was 5PM on February 24, 2026, set by Secretary Pete Hegseth, requiring unrestricted use of Claude for all legal purposes.

Ben New York Times or Washington Post (as cited by Ben)

Sam Altman announced via tweet by 11PM on February 27 that OpenAI had struck a deal with the Pentagon, the same evening Anthropic missed its deadline.

Ben no source cited

The Pentagon government contract signed with OpenAI is worth $200 million per year.

Rick no source cited

Claude was the number one app in the App Store in the week Anthropic refused the Pentagon contract.

Ben no source cited

ChatGPT downloads dropped by approximately 250% in the same period Claude surged to #1 in the App Store.

Ben no source cited

Sam Bankman-Fried invested $500 million in Anthropic in April 2022 using FTX customer funds, receiving a 13.5% equity stake that has since diluted to approximately 8%.

Rick no source cited

SBF's 8% diluted Anthropic stake would be worth approximately $30 billion at Anthropic's current valuation — enough to repay FTX creditors roughly three times over.

Rick no source cited

Anthropic is currently valued at approximately $380 billion and is consistently valued at roughly half of OpenAI's valuation.

Rick no source cited

AI token prices have fallen 850 times since 2020, from $60 per million tokens for GPT-3 DaVinci to 6 cents for open-source Llama 4.

Ben no source cited

OpenAI's frontier model token price dropped from $30 per million in Q1 2023 to $0.50 today; Anthropic's from $15 to $1; Google's from $7 to $0.07.

Rick no source cited

Ben pays $200/month for Claude Max but consumes roughly 500 million tokens per month — equivalent to approximately $12,500 at API pricing.

Ben no source cited

Claude Code accounts for approximately 4% of all GitHub commits and is projected to reach 20% by end of 2025.

Rick no source cited

Icon.com spent $12 million on its domain name, was funded by Peter Thiel and OpenAI, and has apparently gone bankrupt.

Ben no source cited

Higgs Field grew from 15 million users to generating 4.5 million videos per day and is on track for $1 billion in ARR.

Ben no source cited

Cluely founder Roy publicly told TechCrunch the company was doing $7 million ARR when it was actually around $4–5 million, later issuing a public retraction.

Ben TechCrunch

GPT-5.4 scored 75% on the OSWorld computer-use benchmark, up from 47%, surpassing the human baseline.

Ben no source cited

A Mac Mini running a quantised 32-billion-parameter Llama model takes 34 seconds to write a 400-word section, while Claude Sonnet in the cloud completes the same task in 7 seconds.

Rick no source cited

SanDisk's stock rose from a low of approximately $30 to $600 — roughly a 20x gain — within the past year, driven by AI memory demand.

Rick no source cited

Samsung stock is up approximately 250% in the past year due to AI chip and RAM manufacturing demand.

Rick no source cited

Copper is up approximately 20% in the past year and is at an all-time high, breaking into price discovery mode.

Rick no source cited

This episode

Cast

Stats

Episode stats

Insight Overview

insights
chapters

Insight distribution

Sub-Categories

Speaker breakdown

Talk Time

No links parsed

We scan show notes for social handles, websites and apps. Nothing matched on this episode.