▶ YouTube Summaries

• https://www.youtube.com/watch?v=5paRa6E5rCM
• Original title: The Token Black Market

Chinese resellers offer Claude tokens at 70–90% below official Anthropic prices through a "shadow API" gray market. Prime breaks down the mechanics of this economy — how it routes traffic, how it mass- registers accounts despite verification, and the layers of fraud that make the tokens so cheap — then offers his own theory that compromised user machines are quietly part of the network.

The fundamental unit: transit stations

• A "transit station" takes requests from users in regions blocked from OpenAI/Anthropic and proxies them out to the real APIs, piping results back — effectively a VPN, functionally like OpenRouter but illegal.
• These are the building blocks of the whole black/gray market.
• Important framing: it is not primarily a distillation attack. Users are doing real work (programming, building apps, scamming) — real requests and real results.

How they mass-register accounts

From the article "How to buy cheap Claude tokens in China." Three barriers, all defeated cheaply:

• Browser fingerprint: anti-detect headless browsers now mimic real Chrome on the wire — network-layer defenders effectively can't tell legit from fake traffic.
• Phone verification: cheapest defeat in the pipeline — bulk SIMs and SMS numbers cost ~1 cent each. SMS is no longer a proxy for personhood.
• ID verification: a Telegram operator ("John Wick") sells AI-generated licenses/passports/IDs at $15 each, up to 20,000/day. They pass because the MRZ encodes valid check digits per ICAO 9303.
• For real ID + selfie checks, "KYC manufacturers" from the crypto world pay locals (e.g. ~$5 for a smiling photo) and resell IDs and even World Coin eyeball scans.
• The system is modular and replaceable — taking down one SMS provider or ID source doesn't dent it. Practically impossible to stop.

Why the tokens are so cheap — layers of fraud

• Account pooling: route across many Claude Max plans to spread weekly limits — gets cost close to break-even but isn't the main lever.
• Shadow APIs / deceptive model claims: they advertise a frontier model but silently serve a cheaper one. A study found shadow APIs drop Gemini 2.5 Flash accuracy on Med-QA from ~84% (official) to ~37%. You don't know what model you're actually getting.
• Stolen credit cards: pay for Max plans that run free until caught.
• Selling the data: the real prize. Real question/answer pairs from frontier models (Opus, etc.) get sold to Chinese labs to train open-source models that trail ~6 months behind.

Prime's theory: unwitting user machines as mini-VPNs

• Supply-chain worms (Shai-Hulud 1/2/3 in npm, plus compromises in Arch, PyPI) hijack tokens and machines.
• He suspects many "vibe coders" with Claude Code running have their laptops quietly spawning background Claude Code sessions doing work for transit stations — explaining the flood of users hitting weekly limits inexplicably fast.
• Done intermittently and while the user is already active, so it's hard to notice and doesn't get the account banned.
• Ties into the broader mess of bloated, worm-ridden package ecosystems where a "hello world" pulls tens of MB of dependencies.