A 10% Claude proxy works because it borrows someone else's account, farms free credits in bulk and may answer with a cheaper model. The savings are real, the advertised model is not guaranteed, and your code and context go to the operator.
What a resold Claude actually is
A resold Claude endpoint is a relay server, not a genuine Anthropic discount: you point your Anthropic Base URL at the operator's address, the operator forwards the prompt from its own accounts, and the reply comes back stamped as Anthropic. The Anthropic documents that Base URL override as a normal configuration step, and that is the whole integration.
The mechanism has a Chinese industry term that translates roughly as a transfer station. A server sits between you and Anthropic, holds its own credentials, and passes your request through. No VPN, no international card and no identity check are required from the buyer.
That is why the offer spreads through Telegram groups, Mercado Livre listings and GitHub catalogues. The setup cost is one environment variable, so nothing technical stops an operator from signing up customers within an afternoon.
Who runs the relays and how they sell them
Public GitHub catalogues list dozens of these relay stations with converted prices, supported model names and a working status column. The catalogue format is the tell: it looks like any other price comparison table, which makes the service feel like a normal vendor.
The transcript names two public repositories as examples of the ecosystem. One is a catalogue of relay stations, the other a self-hostable relay server published by a developer who wanted to avoid depending on other people's privacy practices. Skalablog has not verified either repository's current contents, licence or listing counts, so treat both as examples of a pattern rather than as reviewed recommendations.
Operators reportedly work in specialist teams: SMS farms for phone verification, sales teams moving accounts in bulk, engineers watching vendor code for changes, and payment staff wiring up local methods. That last group is why some accept Pix, the Brazilian instant payment system run by the Central Bank of Brazil.
How 10% pricing works under the hood
Selling below cost only makes sense if the operator is not paying list price per request. The transcript describes a stack of inputs: bulk account creation to harvest free signup credits, education and enterprise accounts added to the same pool, and a fixed-price subscription carved into token quotas and resold per seat or per hour.
The Chinese phrase for the business model translates roughly as one fish, three dishes. The same underlying access feeds a training pipeline, an internal workload, and a retail channel. Whichever portion is left over becomes the cheap tier you buy.
This is the transcript's account of how the economics are said to work, and it is consistent with the visible behaviour. It is not something an outside party can audit, because the operator controls the accounts, the routing and the logs.
The audit: 45.8% did not return the advertised model
One published audit tested endpoints across resale services and fingerprinted the model behind each reply. Of the endpoints tested, 45.8% returned a model different from the one being sold, according to the study discussed in the video. The claim needs careful handling: the transcript attributes the work to CISPA, the German cybersecurity research centre, but the paper itself has not been opened for this article, so link, title, author list and exact sample size stay unconfirmed.
The method is the useful part. Fingerprinting compares the style, structure and typical failure patterns of a reply against known baselines, much like identifying a speaker by accent. When a service advertises a flagship model and the fingerprint matches a small open-weight model, the substitution is visible to the analyst even when it is invisible to the buyer.
The transcript describes one case where a service selling access to a top-tier model was returning a roughly 9-billion-parameter GLM model. GLM is an open-weight model family from the Chinese lab Zhipu AI.
That substitution changes the calculation. If the delivered model is a cheap open-weight one, a buyer paying 10% of flagship pricing may still be paying more than running that model directly through its own provider.
Why you cannot prove you were served a different model
Resold Claude responses arrive inside a normal reply envelope, so nothing in the payload identifies the model that generated it. The transcript describes operators stamping replies to look like the advertised model. From the client side, a cheap model and a flagship model produce the same JSON shape.
The practical symptom is subtle degradation. Code that needs a correct fix comes back plausible but wrong. You retry, rephrase the prompt, and blame your own prompt craft, because the interface gives you no evidence either way.
Paying users have an additional blind spot. Relays rotate accounts behind the scenes, which drops the cached context attached to the previous session. Operators reportedly compensate by resending the whole conversation on each call, so you keep continuity but pay for more tokens than you would on a direct connection.
The privacy cost: your repository is the product
Every call to a relay sends the operator your prompt, your surrounding context and anything you pasted while debugging. The transcript quotes developers saying the real margin is not in the resale markup but in the logs, because a working engineer's full reasoning trace is expensive training data.
The risk categories are concrete:
- Client source code and internal architecture, sent file by file through the relay.
- Credentials pasted into a prompt during debugging, which are then stored on someone else's disk.
- Commit messages, file paths and personal notes that identify the project and the client.
- Conversation history rich enough to support targeted pressure later, which the transcript's source describes developers reporting.
None of this requires malice at setup. Logging is the default state of any server that terminates your request, and retention is set by the operator.
Account verification, fraud checks and the arms race
Vendors have tightened account verification in response, which raises costs on both sides. The transcript describes Anthropic tripling down on identity checks in April of its own year with document uploads, live selfies and video captured through an outside verification vendor, so that resale accounts fail more often and ordinary users hit more friction.
Anthropic own usage policies bar account sharing and resale of access, and enforcement lands on whatever account is caught, not only on the operator. A buyer whose proxy loses its upstream pool sees an outage they cannot fix, because they hold no account with the provider.
The arms race is documented in resale channels too. The transcript cites a separate analysis cataloguing Telegram channels selling identity bundles with selfie, document and proof of address. Contracting real people to pass checks is the cheaper path, and the transcript points to Brazil's Worldcoin iris-scanning queues in São Paulo as an example of how far that market reached.
Decision guide: what to do instead
The honest answer from the transcript is that the buyer decides, with full knowledge of the risk. The cheaper and safer option is to run the model you actually want, at its real price, through the provider that sells it.
| Option | What you pay | Model accuracy | Code exposure |
|---|---|---|---|
| Direct provider subscription | List price | Advertised model | Provider terms apply |
| Reseller relay at 10% | Fraction of list | 45.8% mismatch risk in the cited audit | Operator sees everything |
| Open-weight model hosted directly | Its own lower rate | Known model | Your own provider |
| Self-hosted relay server | Hardware plus upkeep | Whatever you route | You control logs |
If the goal is running different models inside a coding harness you already like, that is a legitimate workflow and it does not need a relay. Tools and platforms such as Crazystack Typescript and the Bootcamp do Dev Doido teach the configuration side of that setup through supported routes. Gustavo Dev Doido covers related material in his own channels.
Three checks apply before anyone connects a real project. Confirm the endpoint is a first-party API domain. Confirm which account pays for the tokens. Confirm who can read the request logs. A reseller fails all three by design.
Frequently asked questions
- Does a resold Claude use my own Anthropic account? No. The relay holds its own accounts and forwards your prompts, so you never hold a provider relationship and enforcement against the operator takes your access with it.
- How do I check which model answered my prompt? Compare response style and failure patterns against known baseline models, which is the fingerprinting method used in the cited audit. Treat any unexplained degradation in code quality as a signal worth testing.
- Is buying resold access illegal? The transcript treats it as a grey area in contract terms rather than settled criminal law, and breaks provider policy at minimum. Terms-of-service violations can still trigger account closure and unpaid balances.
- What happens to my context when the relay switches accounts? Cached context is lost, and the operator may resend the whole conversation to preserve continuity. That resend increases the tokens billed and hands the operator a longer log.
- Is there any legitimate way to buy access below list price? Provider-run credits, education programmes and enterprise contracts are legitimate routes because the provider sets and honours them. A reseller offering a tenth of list price is not one of those routes.
- Can I run a coding harness with a different model legally? Yes, if you route requests to a provider that sells that model under its own terms and you manage the credentials yourself. Running a relay server on your own infrastructure keeps logs under your control.
- What evidence supports the 45.8% figure? It comes from the audit described in the video, which tested 24 endpoints across three large services. The underlying paper has not been verified for this article, so use the number as a cited claim rather than as an independently confirmed statistic.
- Why do some resellers display a working status column? A price catalogue with availability makes the service look like a normal vendor and shortens the decision. It carries no uptime guarantee, service-level agreement or refund path.
- What is the single largest risk for a working developer? Sending a client's repository through a server that logs it. The model swap hurts one project, while the logged context and credentials can outlive the engagement.
Turning the warning into something you can publish
This story works because a journalist spent a week reading papers, catalogues and forum threads, then explained a hidden supply chain in sixteen minutes of video. The knowledge existed long before the article did; only the format was missing.
If you have recorded an explanation like that, the recording is already the research. Skala Blog turns a YouTube video into a written article by transcribing it, keeping its order and caveats, then shaping it into something a reader can find and cite. Paste the URL, check the draft, publish it.