TL;DR: On June 25, 2026, Greg Kamradt (@GregKamradt, President of ARC Prize) posted that the AI token black market was "obvious in retrospect" — pointing to a Hacker News thread on Anthropic's accusation that Alibaba illicitly extracted Claude capabilities. Commentators described resellers offering Claude at 70–90% below official API prices, Codex at 1–5% of API cost, pooled Claude Max 5x accounts, payment fraud, and reselling reasoning traces to Chinese labs for distillation. For the full Alibaba accusation breakdown (25,000 accounts, 28.8M exchanges), see Anthropic vs Alibaba: distillation at scale.
The HN comment that named the market
The Hacker News discussion on Anthropic vs Alibaba is where the mechanics became concrete. User tristanj summarized how resellers undercut official pricing:
HN comment (June 2026) on item 48667495: resellers offer Claude tokens at 70–90% below official Anthropic API prices via pooled Claude Max 5x accounts, payments fraud, and resale of model output and reasoning chains to Chinese labs.
Core claims from that thread and follow-on posts:
| Mechanism | What it does |
|---|---|
| Pooled Claude Max 5x | Many users share subscription capacity resold as pseudo-API |
| Payments fraud | Stolen or abused cards fund Max seats at scale |
| Proxy gateways | Tools like newapi / sub2api route and meter unofficial traffic |
| Trace resale | Prompts + reasoning chains sold to labs for distillation |
| Free-trial harvesting | Codex/GPT promos in some regions farmed at scale (1–5% of API cost cited on X) |
Greg Kamradt's thread hit 14K+ views in hours — less because the fraud was novel, and more because mainstream builders finally saw how "intelligence too cheap to meter" has a literal gray-market implementation.
Anthropic vs Alibaba: the headline that opened the thread
On June 24, 2026, Reuters and CNBC reported a June 10 letter from Anthropic to Senators Tim Scott and Elizabeth Warren on the Senate Banking Committee.
Anthropic's claims, as reported:
- ~25,000 fraudulent accounts
- 28.8 million exchanges with Claude
- Campaign window: April 22 – June 5, 2026
- Operators affiliated with Alibaba and Alibaba Qwen
- Described as the largest known distillation attack on Anthropic to date
Distillation, in this context, means training a smaller or rival model on outputs (and often chain-of-thought traces) from a stronger model — accelerating capability without matching frontier training spend.
Anthropic had previously flagged DeepSeek, Moonshot, and MiniMax in February 2026 at smaller scales. The Alibaba figures, if accurate, are an order of magnitude larger.
Alibaba did not immediately respond to Reuters' request for comment in published reports.
Why China is ground zero for token arbitrage
Several structural facts show up in every thread — HN, X, and ChinaTalk's prior coverage on cheap Claude tokens:
- Claude and ChatGPT are not officially available in China the way they are in the US — VPNs, foreign cards, and identity friction push users to resellers.
- Official API pricing for Opus-class models is high enough that a 70–93% discount still leaves margin if your cost basis is fraud, pooled Max, or stolen trials.
- Labs compete on price — commentators argue impossibly cheap Claude access forces DeepSeek, GLM, and others to keep consumer API prices low to stay relevant domestically.
That is not just piracy — it is arbitrage on geopolitical access controls.
The Fujian playbook (from X, April 2026)
Months before the HN blow-up, an overheard-in-Fujian story circulated on X (quoted in Greg's thread):
Use hacked credit cards to subscribe to Claude Code Max, resell as API at ~20% of official cost, then sell user traces to Chinese labs for distillation.
Whether every detail is verified, the pattern matches Anthropic's distillation warnings: consumer-facing cheap tokens fund trace harvesting upstream.
Replying developers named stack pieces:
- newapi / sub2api — proxy layers that present one OpenAI-compatible endpoint backed by many upstream accounts
- Regional Codex promos — one-month free trials harvested at scale (@artoriatech cited 1–5% of API cost on X)
Example reseller economics (yunwu.ai)
The HN comment linked yunwu.ai as an example reseller allegedly offering Opus 4.8-class access at roughly a 93% discount versus Anthropic's official API rate card.
We are not linking reseller sites as recommendations — treat them as case studies in pricing pressure:
Official Opus API → list price per million tokens
Gray-market proxy → ~7–30% of list (per HN/X claims)
Reseller margin → subscription fraud + volume + trace side sales
@GregKamradt asked the question builders should ask: "I wonder how often the tokens are cut with other cheap models."
There is no public audit. Ultra-cheap endpoints are a trust problem — you may not get the model you think you are paying for.
Connection to Anthropic ID verification (July 8)
The token gray market intersects directly with Anthropic's July 8, 2026 identity verification policy — government ID, biometrics, facial geometry.
Official rationale: security and abuse prevention.
HN interpretation: tens of thousands of bot accounts used to farm Claude for distillation and reseller pools — the "onslaught" that made signup-based fraud scalable.
Same week as the Alibaba letter, the US also moved on export controls for Fable 5 and Mythos 5 — a different lever, same war: stop frontier capability from leaking to adversarial distillation pipelines.
Who else is in this market?
Beyond end-user resellers in China, @WesamMikhail claimed on X that some YC companies were looking to sell hundreds of thousands of dollars of token capacity — suggesting the gray market is not only overseas proxy shops.
@tynervp pushed back that scale may be overstated. The honest synthesis: documented distillation campaigns are millions of exchanges; consumer reseller pricing is widespread anecdotally; exact GMV is unknown.
What is not disputed:
- Anthropic published account-farm numbers in a Senate letter
- Reseller pricing 70–93% off list is repeatedly cited by developers who monitor CN API markets
- Proxy tooling (newapi, sub2api) is openly discussed in builder communities
What this means if you build with Claude
For individual developers
- Ultra-cheap API keys from Telegram ads or CN marketplaces are high-risk: ToS violations, data exfiltration, model substitution, sudden cutoffs.
- If price is the issue, compare official tiers, batch API, smaller models, or open-weight alternatives before gray-market proxies.
For AI labs and enterprises
- Distillation is an industrial threat — not a hobbyist jailbreak. Rate limits, device binding, and ID checks are economic defenses, not just UX annoyances.
- Reasoning-trace logging is valuable IP. Treat chain-of-thought outputs as sensitive if your model is frontier-class.
For policy readers
- The White House distillation memo (April 2026) pledged coordination against industrial-scale extraction. Anthropic alleges Alibaba ignored those warnings — whether that holds legally is for investigators; the incentive structure is clear.
Agentic stacks sit on top of token economics
Most coding agents — Claude Code, Cursor, Codex — bill through the same upstream APIs resellers attack. When Eric Xing's Critique of Agent Model distinguishes agentic harnesses from agentive autonomy, the black market is a reminder: the harness is only as honest as the token supply behind it.
Cheap tokens enable more agent loops. They also fund trace harvesting that compresses frontier lead times for competitors.
Timeline
| Date | Event |
|---|---|
| Feb 2026 | Anthropic flags DeepSeek, Moonshot, MiniMax distillation campaigns |
| Apr 22 – Jun 5, 2026 | Alleged Alibaba/Qwen campaign — 28.8M exchanges (per Anthropic letter) |
| Jun 10, 2026 | Anthropic letter to Senate Banking Committee |
| Jun 24, 2026 | Reuters / CNBC publish Alibaba allegations; HN thread explodes |
| Jun 25, 2026 | @GregKamradt viral thread — "black market for tokens" |
| Jul 8, 2026 | Anthropic ID verification policy effective (scheduled) |
Related reading
- Anthropic vs Alibaba — 25K fake accounts, 28.8M exchanges
- US export controls on Fable 5 and Mythos 5
- Eric Xing: agentic vs agentive AI
- Why AI companies want you using agents
- Claude Code hooks and approval flows
Primary sources: Hacker News #48667495 · Greg Kamradt on X · Reuters on Anthropic-Alibaba · CNBC · arXiv:2606.23991 (agency framing)
Reseller pricing claims come from HN/X community reports — not Anthropic official pricing. Alibaba has not publicly confirmed or denied the June 2026 allegations as of publication. Verify against primary reporting before citing in compliance docs.