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

title	description	status	category	tags	draft	created	updated
Analysis — LLM Subscription-Resale Markets	Mechanism deep-dive for the InferHub case — supply tiers, credential pipeline, quota windows, egress pooling, auction pricing, measured volume, reliability and risk surface.	active	market	inference-market subscription-arbitrage free-tier-farming credential-resale supply-chain tos-risk	true	2026-10-05	2026-10-05

Analysis — LLM Subscription-Resale Markets

Part of llm-subscription-resale-markets/index

All captures unauthenticated and public. Claims labelled VERIFIED (observed in a capture whose URL and UTC timestamp are recorded), INFERRED (reasoned — reasoning given), UNKNOWN.

1. What the platform actually is

VERIFIED — inferhub.dev/terms §2 (last updated 2026-10-02):

"The Service is 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). Publishers attach their own authorized upstream subscriptions; consumers call the Service through OpenAI- and Anthropic-compatible endpoints using an API key."

It is a two-sided market: the platform owns no supply. Publishers list credentials; consumers deposit USDC and call an OpenAI/Anthropic-compatible API; the platform routes and takes 20%.

VERIFIED — GET /api/pricing/config (public):

{"maxAskPctOfOfficial":0.5,"platformFeePct":0.2,"publisherSharePct":80}

2. Supply: all 14 upstream families are subscriptions

VERIFIED — GET /api/market (public, 200) carries a family per model. Complete list, with the model counts seen on the pricing page:

prefix	family	nature
cc	Claude Code	vendor subscription (OAuth)
cx	OpenAI Codex	vendor subscription (OAuth)
zai	Z.AI GLM Coding Plan	third-party coding plan
ocg	OpenCode Go	third-party coding plan
cmc	Command Code	API-key-based plan
cp	ClinePass	third-party coding plan
cb	CodeBuddy	coding-agent credits (Tencent)
cbcn	CodeBuddy CN	coding-agent credits
ag	Antigravity	Google free/IDE tier
ali / alicn	Qwencloud/Alibaba (+CN)	cloud coding tier
mm / mmcn	MiniMax (+CN)	coding tier
mimo	Xiaomi MiMo	coding tier

Not one of them is a metered frontier API. This is the single most diagnostic fact about the business: the gateway's supply is other people's flat-rate products.

3. Supply tier A — paid subscription arbitrage

The arithmetic that produces the owner's headline number:

cc/claude-opus-5-5 in the market feed:

{"slug":"cc/claude-opus-5-5","family":"Claude Code","minAskIn":0.796,"minAskOut":3.98,
 "maxAskIn":2,"maxAskOut":10,"lastRate":0}

• Official reference: $4 /Mtok input.
• Cheapest ask: $0.796/Mtok = 80.1% off.
• Highest ask: $2.00 = exactly the 50% cap. The cap is a ceiling, not a floor.
• lastRate: 0 — the most recent settled rate on that slug was zero.

INFERRED (strong): with a flat subscription already paid, a publisher's marginal cost per token is ~0, so any positive price is profit; the auction routes to the cheapest ask, so the rational publisher undercuts toward zero. The "80% discount" is therefore the residual of a dumping auction, not a purchase discount. It cannot be matched by any wholesale API deal, because the competitor's true cost is a sunk fee rather than a per-token price.

4. Supply tier B — free-tier farming (the 99% rows)

VERIFIED — InferHub's own public OpenAPI (GET /api/openapi.json, 138,946 bytes, 73 paths) documents bulk credential onboarding as a product feature. Quoted descriptions:

• POST /publisher/providers/antigravity/bulk — "Bulk-add Antigravity (Google) accounts from refresh tokens. The body is the accounts payload itself… Each item needs a refreshToken (an Antigravity access token lives about an hour, so the refresh token is the durable input); email is optional and only names the row. Each account is refreshed for a live access token and onboarded onto a Cloud Code Assist project before it is stored — accounts Google will not onboard are reported as failures rather than saved."
• POST /publisher/providers/codex/bulk — same shape for Codex.
• POST /publisher/providers/codebuddy/refresh-tokens — bulk OAuth refresh for CodeBuddy.
• POST /publisher/providers/recheck-bulk, GET /publisher/providers/invalid — fleet-wide credential health sweeps. The platform expects credentials to die routinely.
• GET /publisher/providers — always paginated because "a single publisher can own six figures of credentials."
• PATCH /publisher/providers/{id}/auto-reset-credit — "Toggle Codex auto reset-credit spend (Codex only; auto-spends real money)."
• GET /publisher/upstreams/{slug}/max-concurrent, .../schedule — per-upstream concurrency and scheduling.

4.1 The vendor free tier being farmed

VERIFIED — CodeBuddy (Tencent / "Buddy AI" international site), from its own pricing and subscription pages:

Plan	Price	Credits
Free	$0	100 credits/month
Free daily activity bonus (limited-time)	$0	+30 credits/day
Free, new user	$0	250 credits on first sign-in, valid 14 days
Pro	$10/mo ($96/yr)	2,000/mo (1,000 base + 1,000 bonus), +50/day
Pro 7-day trial	card + auto-renew	500 credits
Team	$40/seat/mo	1,000/seat/mo

While the promotion runs, "all models are available, code completion is unlimited, and up to 99 automated tasks are included." Base and bonus credits do not roll over month to month, and the daily bonus is "non-transferable, non-giftable, and cannot be exchanged for cash".

⇒ A fresh free account yields roughly 1,000 credits/month at zero cost.

VERIFIED — the vendor bans the method, in its own words (Subscription Policy, "Activity Restrictions" — rewards may be cancelled, reclaimed, and eligibility suspended):

"1. Participating using false, invalid, or non-personal account information; 2. Using scripts, plug-ins, emulators, API calls, bulk registration, or other means to farm activity or rewards; 3. Exploiting system vulnerabilities, network abnormalities, or technical faults for improper gain; 4. Maliciously creating or controlling multiple accounts to participate; 5. Violating Buddy AI's terms of service, subscription rules, credit usage rules, or applicable laws and regulations; …"

and in the FAQ: "…users may not obtain improper rewards through bulk registration, account sharing, account trading, or other abnormal methods."

VERIFIED — geographic control: www.codebuddy.ai does not resolve from US, Indian or Russian networks (DNS answers 0.0.0.1 to client subnets there). The resale market routes around it.

4.2 The credential market that supplies it

VERIFIED — mastumbas.id/product/codebuddy, an Indonesian digital-goods marketplace (QRIS / local rails, "Made in Indonesia"):

• "Codebuddy Global 480-500++Credit [BERGARANSI] (Auto Claim Harian +30 Credit Daily)", store Dudul, listed Rp 123,456 → Rp 2,300 (≈ $0.14), 3,303 sold, stock 720, 2-day warranty, 48 reviews at 5.0.
• Wholesale: ≥10 Rp 2,200 · ≥20 Rp 2,100 · ≥30 Rp 2,000 · ≥50 Rp 1,900 · ≥100 Rp 1,800/pc. (100 accounts ≈ Rp 180,000 ≈ $11.)
• Seller's description: "1 token = 1 CodeBuddy account containing 450 credits, usable in 9router for 28 AI models: text, coding, image and video. Credits increase +30 per day automatically as long as the account is still active." Delivery = access token + refresh token as a .json.
• Adjacent listings on the same page: Cline accounts (Rp 2,000, "limit per model 10–15 M tokens daily, auto reset 24 h") → cp/; "Refresh Token Codebuddy CN 4000-5000 Credit" (Rp 15,000) → cbcn/; "Apikey Codebuddy Global 30000 Credit" (Rp 100,000); a CodeBuddy-CN listing attributed to rdhx (the operator named in the earlier free-token-gateway study); Gemini Pro links; and 1-day Webshare proxy packs. Account supply and proxy supply on the same shelf.

VERIFIED (set comparison) — InferHub cb/ slugs vs the 28 model codes that listing advertises: 19 exact matches of InferHub's 23.

claude-opus-4.6  claude-opus-4.7-1m  deepseek-v4.1-flash  gemini-2.5-flash-image
gemini-3.1-flash-image  gemini-3.1-pro  glm-5.2  glm-5.3  gpt-5.3-codex  gpt-5.4  gpt-5.5
gpt-5.6-luna  gpt-6-astra  gpt-6-sol  gpt-image-2  grok-4.6  grok-4.7  hy4-preview  kimi-k3

InferHub-only: claude-opus-5, gemini-3.0-pro-image, gpt-6-luna, minimax-m3. Listing-only: gemini-3.5-flash, gemini-3.8-flash, glm-5.3-flash, gpt-5.6-sol, gpt-5.6-terra, hy3, kimi-k2.6, kimi-k2.8-preview, seedance-2.5.

⇒ InferHub asks $0.04/Mtok for cb/claude-opus-4.6 and cb/claude-opus-4.7-1m — 99.2% off an official $5 — and that capacity is reachable through a ~$0.14 account from this market.

INFERRED (not asserted as fact): that MasTumbas supply specifically feeds InferHub. What is verified is that InferHub names the family "CodeBuddy", ingests bulk CodeBuddy refresh tokens, and serves the same model-code set; MasTumbas is a public example of the supply, not proven to be the identical accounts.

5. Routing and failure semantics

VERIFIED — inferhub.dev/docs, "Routing tips":

• "Auto routing: The router picks the best publisher serving that prefix per request, falling over to another on failure (up to 5 hops)."
• 429 capacity_saturated (Anthropic envelope rate_limit_error, Retry-After: 2) when every provider for a model is at its concurrency limit. Nothing billed.
• "Quotas: If every publisher serving a prefix has hit their plan's monthly cap, we return 503 no_capacity until any of their periods rolls over."
• 402 no_provider_under_bid with min_input_per_mtok / min_output_per_mtok when no provider meets a consumer's bid headers.
• x-ignore-upstream: cmc,qd,cx — a prefix exclusion list that also reveals an unlisted qd (Qoder) prefix.
• Model addressing: <prefix>/<model> (e.g. ocg/glm-5.2), a bare alias (cheapest-first across members), combo/<slug>, or free/<slug>/<model>.
• Minimum charge $0.00001 per billed request; failed requests, no-output requests and free models cost $0.

Quota modelling: the home page states "We poll Claude, Codex, and z.ai for live 5-hour and weekly usage windows. No manual caps — exhaustion is detected from the source of truth." Together with the usage-windows provider sub-resource and auto-reset-credit, this reads as: the platform authenticates to the vendor's own usage/limit API with the publisher's OAuth token and tracks each credential's 5-hour and weekly windows, skipping exhausted ones. INFERRED for the internal detail; VERIFIED that quota windows are polled and surfaced.

6. The egress layer

VERIFIED — platform copy: "Per-provider outbound proxy pool routes your traffic through least-used egress."

VERIFIED — the owner-supplied Discord screenshot (owner-proxy.png) is a proxy-vendor dashboard: "1,000 Proxies", "∞ GB bandwidth/month", start "24 Sep 2026 16:23", "2,866,045 Total Requests", "4.4 TB / ∞ GB Total Bandwidth used". Posted by handle ExZork answering "penasaran sama bandwith usage proxinya inferhub ngehandle puluhan ribu akun" — "curious about the proxy's bandwidth, InferHub handles tens of thousands of accounts."

INFERRED: spreading many accounts across many egress paths prevents both rate-limit correlation and the vendor's own proxy/VPN detection (Anthropic is documented as detecting datacenter-range traffic). The same marketplaces that sell the accounts also sell the proxies.

7. Measured scale

VERIFIED — GET /api/pricing/usage (public), window 2026-10-04T19:30Z → 2026-10-05T19:37Z: 152 routable slugs, 5,684,357 requests, 406,505,881,779 input tokens (333.3 bn cached = 82.0%), 9,616,677,225 output tokens.

family	requests	input tokens	tier
CodeBuddy (cb/)	2,843,337	266.8 bn	free-tier credits
CodeBuddy CN (cbcn/)	1,972,650	92.1 bn	free-tier credits
OpenAI Codex (cx/)	343,683	15.9 bn	subscription
Antigravity (ag/)	156,661	6.1 bn	Google free/IDE tier
Qwencloud/Alibaba (ali/)	155,240	10.9 bn	cloud coding tier
Qwencloud/Alibaba CN (alicn/)	93,010	5.0 bn	cloud coding tier
Z.AI GLM (zai/)	48,603	4.1 bn	subscription
Claude Code (cc/)	32,249	2.77 bn	subscription
MiniMax (mm/)	21,573	1.57 bn	coding tier
Command Code (cmc/)	7,430	0.28 bn	subscription
OpenCode Go (ocg/)	5,990	0.64 bn	subscription
ClinePass (cp/)	3,636	0.30 bn	subscription
Xiaomi MiMo (mimo/)	181	3.0 M	coding tier
MiniMax CN (mmcn/)	114	6.0 K	coding tier

CodeBuddy alone is 50% of requests and 66% of input tokens. Claude Code is 0.57%. The flagship-discount rows are marketing surface; the throughput is farmed free-tier commodity capacity. Largest single slug: cb/deepseek-v4.1-flash, 2,434,748 requests, 242 bn input tokens.

VERIFIED — pricing page headline counts: 14 curated upstreams, 155 routable model IDs, 74,310 active providers (marketplace supply). One Claude slug (claude-opus-4.7) shows 6,758 available slots.

8. Reliability and latency

VERIFIED — GET /api/status (public): overall 24 h availability 99.6%, 7 d 99.59%; API availability 24 h 96.4%. API error causes: no_capacity 62.7%, upstream_failed 37.2%, billing_unavailable 0%.

Auto-route alias availability: claude-fable-5 35.07% over 24 h (currently degraded at 82.35%); claude-fable-5.1 major_outage, 27.5% current.

prefix	family	avail 24 h	p50	TTFT 1 h	tok/s
cp	ClinePass	93.10%	7.5 s	5.0 s	62.8
alicn	Qwencloud/Alibaba CN	96.72%	30 s	18.1 s	27.2
cx	OpenAI Codex	96.84%	15 s	8.5 s	23.6
cc	Claude Code	97.71%	5 s	5.1 s	42.0
cmc	Command Code	98.30%	60 s	8.6 s	39.6
ali	Qwencloud/Alibaba	98.90%	7.5 s	5.0 s	41.7
ocg	OpenCode Go	99.48%	15 s	8.4 s	24.7
zai	Z.AI GLM Coding Plan	99.50% (degraded)	15 s	18.7 s	24.3
ag	Antigravity	99.78%	7.5 s	10.6 s	73.9
mm	MiniMax	99.84%	7.5 s	4.6 s	38.2
cb	CodeBuddy	99.85%	7.5 s	11.5 s	115.2
cbcn	CodeBuddy CN	99.98%	10 s	5.9 s	39.0

First-token times of 5–19 s and a 3–7% failed-request rate are not first-party API characteristics. INFERRED: the discount is paid for in availability and latency, and a capacity-constrained pool is the binding constraint (no_capacity dominating the error mix).

9. Risk surface

VERIFIED (policy text) — InferHub privacy policy:

• On an upstream security-policy rejection, the system prompt, tool definitions and request parameters of that request are retained 7 days ("the scaffolding of that one request").
• Safety screening reads the consumer's messages and system prompt against a keyword/pattern list before sending upstream, expressly "to keep content prohibited by our Acceptable Use terms (material involving minors, and malicious tooling) off the provider accounts that serve you" — i.e. the pooled provider accounts are the asset being protected. Flagged text is retained 14 days, rule/score metadata 90 days. Tool results, file contents and attachments are excluded.
• Publisher credentials (API keys, OAuth refresh tokens, outbound proxy credentials) are stored AES-256-GCM encrypted (api_key_enc, oauth_refresh_enc) and decrypted "only in-process, in memory, when a request is routed".

VERIFIED (docs): some upstreams "add to or reshape your system prompt (for example, Cursor prepends its own agent prompt)" — surfaced as systemPromptNote from GET /api/catalog. A publisher's agent scaffolding can therefore land inside a consumer's request.

INFERRED: (a) plaintext credentials exist in the operator's process memory, so a host compromise or a rogue operator exposes live vendor credentials for every connected subscription; (b) consumers cannot verify which credential served them, so a substituted backend is not detectable from the outside.

VERIFIED — Terms of Service posture: §7 puts every upstream-ToS obligation on the publisher ("represent and warrant" authority; "solely responsible" for compliance "including any restrictions on shared, transferred, or resold access"; indemnify InferHub). §8: balances are non-refundable and a chargeback forfeits the whole balance. §6: paid capacity may be resold, free-pool capacity may not. §10: no warranty on output rights.

10. Legal exposure (the landing vendor's own words)

VERIFIED — code.claude.com/docs/en/legal-and-compliance:

"Customers may not pay for, resell, or intermediate Claude usage on their end users' behalf. Each end user must authenticate with their own Anthropic API key, Claude subscription plan credentials, or 3P inference provider credential…"

"Anthropic does not permit third-party developers to offer Claude.ai login into their own applications, or to route requests through Free, Pro, or Max plan credentials on behalf of their users. Moreover, developers may not collect, store, or intermediate Claude.ai credentials or session tokens — sign-in to a Claude account must complete through Anthropic's own flow."

For the Claude Code upstream, InferHub does all three at once. INFERRED: the platform's indemnity chain (publisher warrants authority; publisher indemnifies) cannot extinguish the upstream's own prohibition — it only reallocates who pays. Secondary sources report enforcement tightening through 2026 (subscription tokens blocked in third-party harnesses), which is consistent with the platform's need for 1,000 rotating proxies and its invalid / recheck-bulk sweep endpoints. UNKNOWN: whether any specific publisher or consumer has been banned (deep lane pending).

11. Model-substitution risk

INFERRED, literature-supported. The same market is the subject of black-box auditing work: "Which Model Is Actually Serving You? IRIS: Budgeted Black-Box Auditing of Model Substitution and Routing Dilution in LLM Gateways" (arXiv 2607.20860) and "Your 'Pro' LLM Subscription May Actually Be 'Free': Exposing Fingerprint Spoofing Risks in LLM Inference Services" (arXiv 2606.16100). Both treat "the served backend may not match the advertised one" as a known property of commercial gateways. Combined with settling prices at 0.1–1% of official on free-tier families, substitution or mislabelling cannot be excluded from public data. UNKNOWN for InferHub specifically — settling it needs a paid account plus a fingerprint audit, which is outside the passive scope.

12. This is a category, not an original

• Surplus Intelligence (surplusintelligence.ai) — "Open Market for Models: your API calls routed to the cheapest offer, your surplus capacity sold to the highest bidder", with "list your API credits" as the seller path; Base-native with a SURPLUS token (116.5 bn LLM tokens reported through that ecosystem per a public project page).
• inferencehub.tech — same niche; its own comparison copy places "Gray-market supply · Start in minutes" opposite "Real supply through hyperscaler and enterprise channels", and states plainly "We're not cheaper than the labs."
• FreeLLMAPI — aggregates 34 free providers / 635 endpoints, claims 7.4 bn tokens/month.
• OmniRoute — peer gateway that publishes a ToS table naming 15 providers whose terms prohibit proxy access, i.e. the community documents its own prohibited-use footing.
• The September 2026 study (RESEARCH/free-token-gateways/) found four separate operations doing the operator-run version of the same thing, with first-party admission of bulk account creation, one-account-per-proxy signup, and edu-mail resale inside the operator's own community.

Conclusion of the analysis: the cheap-frontier-model product is a credential logistics business dressed as an API company. Its cost base is other people's flat-rate entitlements and free tiers; its moat is the operational machinery (bulk onboarding, token refresh, quota polling, egress rotation, failover) rather than any supply relationship; and its main risk is that the landing vendors actively prohibit the arrangement.