research/llm-subscription-resale-markets/index.md

---

title: LLM Subscription-Resale Markets description: How "one API key, every frontier model" gateways sell Claude/GPT at 80–99.9% off — subscription-entitlement resale, free-tier account farming, proxy-pooled credential fleets, and the auction that prices it. Primary case: InferHub (inferhub.dev). status: active category: market tags: [inference-market, subscription-arbitrage, gray-market, free-tier-farming, credential-resale, supply-chain, tos-risk, codebuddy, inferhub, usdc] draft: true created: 2026-10-05 updated: 2026-10-05 origin: buzz://d8a718be-031f-4a6e-9f5c-a55466641654/51d7492cdba77ffc37bde707ab87d950a450debdb2ab50d288ffec778d6f93bd

LLM Subscription-Resale Markets

Summary

A class of inference gateways sells frontier-model access far below the vendor's list price (80% off on a Claude subscription family, up to 99.9% off on others). They do not buy tokens from the model vendors and they do not have a wholesale deal. They operate a two-sided market for other people's subscription entitlements and free-tier credits, priced by a cheapest-wins auction, and they route the traffic through pooled credentials fronted by rotating egress proxies.

Primary case studied: InferHub (inferhub.dev), a USDC-on-Solana marketplace with 74,310 listed active providers, 155 routable model IDs, and 5.68 M requests / 406.5 bn input tokens measured over one 24-hour window.

Research Question

How can a third-party gateway serve frontier Claude/GPT/Gemini models at 80% or more below the vendor's official per-token price — and what does that architecture imply for the people buying and selling through it?

Scope

Included: the supply side (where the capacity comes from), the routing/auction mechanism, the credential-ingestion pipeline, the measured scale, the reliability profile, and the terms-of-service/legal exposure for both sides of the market.

Excluded: attempting to authenticate, testing any credential, exploiting any endpoint, and black-box model-fingerprinting (needs a paid account — left as a next step).

Sufficient answer: the supply mechanism is identified and its chain traced to a concrete, purchasable artefact, with the decisive claims backed by first-party or vendor-primary sources.

Method: passive, unauthenticated reconnaissance only. Public pages, public JS/CSS, public API schemas and unauthenticated error responses, public marketplaces, DNS, npm, public code hosts. No account created, no key sent, no auth bypassed, ~1 s pacing.

Key Findings

• The model is subscription resale, admitted in the operator's own terms. InferHub's ToS §2: "a reverse proxy and marketplace that resells access to upstream LLM subscription entitlements, including subscriptions accessed via OAuth (such as Claude and Codex), API-key-based plans (such as Command Code), and coding plans offered by third parties (such as z.ai GLM Coding Plan and OpenCode Go)."
• All 14 upstream families are consumer coding-agent subscriptions, not metered APIs. GET /api/market (public) returns a family per model: CodeBuddy, CodeBuddy CN, Command Code, OpenCode Go, ClinePass, Claude Code, OpenAI Codex, Qwencloud/Alibaba (+CN), Antigravity (Google), MiniMax (+CN), Z.AI GLM Coding Plan, Xiaomi MiMo.
• The discount is auction clearing price, not a purchase discount. GET /api/pricing/config returns {"maxAskPctOfOfficial":0.5,"platformFeePct":0.2,"publisherSharePct":80}: a publisher ask is capped at 50% of official, and competition between publishers pushes it far below the cap, because the seller's marginal cost is approximately zero once the flat subscription is paid.
• There are two distinct supply tiers under one roof, and they are different businesses:
  • Paid subscription arbitrage (cc/ Claude Code, cx/ Codex, zai/, ocg/ OpenCode Go, cmc/ Command Code, cp/ ClinePass) — a real sunk monthly cost per account. This is the tier that produces the owner's headline 80% figure: cc/claude-opus-5-5 asks $0.796/Mtok against official $4.
  • Free-tier / promo-credit farming (cb/ CodeBuddy, cbcn/, ag/ Google Antigravity, ali*/) — acquisition cost is one account signup, and on the secondary market ~$0.11–0.14 per account. This is the tier that produces the 99% rows.
• Free-tier farming is productised as a platform feature. InferHub's public OpenAPI documents POST /publisher/providers/antigravity/bulk — "Bulk-add Antigravity (Google) accounts from refresh tokens … each account is refreshed for a live access token and onboarded onto a Cloud Code Assist project" — plus */codex/bulk, */codebuddy/refresh-tokens, */recheck-bulk, */invalid, */usage-windows and */auto-reset-credit ("Codex only; auto-spends real money"). A single publisher is documented as able to hold "six figures of credentials".
• A concrete, purchasable artefact sits at the bottom of the chain. Indonesian marketplace mastumbas.id sells CodeBuddy accounts — access token + refresh token, 450 starting credits, +30 credits/day auto-claim — at Rp 2,300 (~$0.14) retail, Rp 1,800 at ≥100 units, 3,303 sold. Cline accounts, CodeBuddy-CN refresh tokens and 1-day Webshare proxy packs sell on the same shelf.
• Decisive link (slug correspondence, VERIFIED): 19 of InferHub's 23 cb/ slugs match that listing's advertised model codes exactly, including claude-opus-4.6 and claude-opus-4.7-1m — which InferHub asks $0.04/Mtok for against an official $5, i.e. 99.2% off. That capacity is reachable through a $0.14 account, not through any arrangement with Anthropic.
• The vendor's free tier is the engine: ~1,000 credits/month at zero cost. CodeBuddy Free = 100 credits/month + 30 credits/day (limited-time daily-activity bonus) + 250 welcome credits; all models available while the promotion runs. Tencent bans exactly this in its own subscription policy ("scripts, plug-ins, emulators, API calls, bulk registration, or other means to farm activity or rewards"; "maliciously creating or controlling multiple accounts"; "bulk registration, account sharing, account trading").
• The volume is in the cheap commodity tier, not the flagships. CodeBuddy is 50% of all requests and 66% of all input tokens; the real Claude Code family is 0.57% of requests. The 80%-off Claude row is the shop window.
• The landing terms forbid the business model outright. Claude Code's own legal & compliance page: "Customers may not pay for, resell, or intermediate Claude usage on their end users' behalf" and developers "may not collect, store, or intermediate Claude.ai credentials or session tokens". InferHub does all three for the Claude Code upstream.
• Consumers pay for the cheapness in failure rate and latency, and in data exposure. API availability 24 h 96.4%, with error causes no_capacity 62.7% / upstream_failed 37.2%; TTFT 5–19 s; claude-fable-5 auto-route at 35% 24 h availability. InferHub's privacy policy retains the consumer's system prompt and tool definitions for 7 days on security-policy rejections, retains screened message text for 14 days, and documents that some upstreams prepend their own agent prompt to the consumer's request.
• The mechanic is a category, not an original. Surplus Intelligence runs the same open market on Base rails ("your surplus capacity, sold to the highest bidder"); inferencehub.tech in the same niche labels the sector "gray-market supply" in its own comparison copy; FreeLLMAPI aggregates 34 free providers; OmniRoute publishes a ToS table of 15 providers whose terms prohibit proxy access.

Sources

Full reference list with per-source context: llm-subscription-resale-markets/sources.

Decisive ones: InferHub /terms, /privacy, /docs, /faq, /api/openapi.json, /api/pricing/config, /api/market, /api/pricing/usage, /api/status; code.claude.com/docs/en/legal-and-compliance; codebuddy.ai pricing/subscription docs; mastumbas.id/product/codebuddy; npm @inferhub/helper.

Detail Files

• llm-subscription-resale-markets/analysis — the mechanism in full: supply tiers, the auction, the credential pipeline, quota windows, egress pooling, measured volume, reliability, risk surface
• llm-subscription-resale-markets/sources — every reference, with what it evidenced
• llm-subscription-resale-markets/next-steps — open questions and the evidence that would settle them

Related Research

• Prior study, workspace only: RESEARCH/free-token-gateways/ (2026-09-15) — the operator-run version of the same supply (pooled CodeBuddy/free-tier accounts, bulk email + proxies + stealth browser, custom gateway; LiteLLM falsified). Four separate operations, no shared operator.
• Workspace raw material for this topic: RESEARCH/inferhub-marketplace/ — ORCHESTRATOR_INVESTIGATION.md, SUPPLY_CHAIN_CODEBUDDY.md, RECON_NOTES.md, raw captures under .scratch/EVIDENCE/.

Discussion

Raised by brutaljokerz in #research on 2026-10-05 while looking at inferhub.dev/pricing ("how does it serve models very cheap especially opus 5.5, frontier at 80% discount?"), with an operator-posted screenshot of a proxy-vendor dashboard (1,000 proxies, 4.44 TB, 2.87 M requests since 24 Sep 2026) posted by Discord handle ExZork in reply to a question about the platform "handling tens of thousands of accounts".

Follow-up in the same thread connected it to the September free-token-gateway study and to the CodeBuddy-account resale market; the operator's npm package @inferhub/helper carries maintainer exzork, which matches the Discord handle. The platform describes itself as "operated by its developer as an individual"; the jurisdiction is not disclosed.

Operator screenshot retained at RESEARCH/inferhub-marketplace/.scratch/EVIDENCE/owner-proxy.png. Thread: buzz://d8a718be-031f-4a6e-9f5c-a55466641654/51d7492cdba77ffc37bde707ab87d950a450debdb2ab50d288ffec778d6f93bd

Next Steps

• Fold in the three deep lanes still running at time of writing (credential/quota mechanism, GMV economics with the two-tier split, and the remaining 13 vendors' upstream clauses)
• Owner to add additional material in-thread
• Decide draft: false if this goes to the public Quartz site

Conclusion

Open — status: active. The supply mechanism is settled; the economics quantification and the enforcement record are not.