· 4 min read · Engineering
Your cheap Claude is probably a fake. We caught 183 of them.
We probed 8 popular Claude resellers for 17 days. 183 of their channels were not Claude at all. Most were Kiro Cascade or Codeium wearing a Claude name tag. Names, numbers, and the script we used so you can test your own provider.
By Inferio team
- engineering
- announcement
How the probes work
What is not spoofing: Bedrock, Vertex, Foundry
What 17 days of probing turned up
Why this happens
Test your own provider in 5 minutes
Claude is the most in-demand model on the market right now, and one of the most expensive. That gap has created a thriving market of third-party resellers offering "the same Claude" at a fraction of the official price. Some are legitimate. A lot aren't.
Over 17 days of automated probing across 8 upstream resellers, we found 183 (channel, model) pairs that fail authenticity checks against models marketed as claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5, and friends. The same pressure exists for every premium model, so GPT, Gemini, and Grok are almost certainly affected too. Claude is just where the price gap is biggest right now and where we have hard data.
Per (channel, model) we send four parallel requests to the standard Anthropic Messages API (POST /v1/messages):
Probe
Prompt
Pass criteria
Tell me a 2-sentence sad story about a lost kitten.
No coding-refusal phrases; response >= 40 chars
Write a haiku about the ocean at sunrise. Only the haiku, nothing else.
No coding-refusal phrases; response >= 25 chars
What company created you? Reply with only the company name, one word.
Contains anthropic; no foreign-vendor names
Which model are you? Reply with only your model name, nothing else.
Contains claude or anthropic; no foreign-vendor names
Three failure types:
coding-tool-refusal: response matches a Kiro Cascade or Codeium / Windsurf Droid refusal pattern ("I'm here to help with coding", "I'm Droid", etc.). The upstream is routing to an IDE-assistant product, which refuses the prompt because it's not coding.
foreign-identity: response identifies as a non-Anthropic vendor (OpenAI, Meta, DeepSeek, Moonshot, etc.).
failed: wrong-style output without those signals (no kitten story, no haiku, generic "AI assistant" reply).
Pattern lists and probe code: inferionet/inferio-frontend.
A reseller routing through AWS Bedrock, Google Vertex AI, or Azure AI Foundry is not spoofing. Those platforms host real Anthropic Claude weights under license. Buying capacity there at enterprise discount and reselling it is a normal supply chain.
One known false positive: cloud-hosted Claude sometimes answers the identity probe with Amazon / Google / Microsoft instead of Anthropic, because of host system prompts. A channel flagged only by foreign-identity against cloud-host names warrants a manual second look. coding-tool-refusal and non-cloud foreign vendors are unambiguous, real Bedrock/Vertex/Foundry Claude never produces those responses.
8 upstream resellers (anonymized as provider-1..provider-8), 183 (channel, model) entries between 2026-04-08 and 2026-04-24.
Failure type
Count
Probe labels that triggered failures (a channel can fail multiple):
Failures
emotional being the biggest catcher is the smoking gun. Real Claude doesn't refuse to write a 2-sentence sad story about a kitten. A coding-tool backend dressed as Claude does, every time.
Provider
Bad channels
Model
Older Claude variants have higher counts only because they've been on the market longer. Opus 4.7 hit 16 spoofed channels within weeks of release. This isn't a static snapshot, either, resellers rotate upstreams and a channel that passes today can start serving Kiro tomorrow. Authenticity has to be checked continuously.
Real Claude is expensive whether you buy from Anthropic or from a licensed cloud reseller. A reseller advertising Claude below the cheapest licensed price has four options:
Eat the loss (sustainable only with deep funding).
Buy Bedrock/Vertex/Foundry capacity at enterprise discount and resell. Legitimate, but bounded by what those clouds charge.
Run a fraction of traffic through real Claude and route the rest cheaper, banking on users not noticing.
Route everything to a non-Anthropic backend and hope nobody checks.
Options 3 and 4 are what the probes catch. Kiro Cascade and Codeium are tempting backends because they have free / near-free quotas and Anthropic-compatible response shapes. The output looks structurally correct, just stylistically wrong, and most users never notice unless they ask for something non-coding.
We built these probes into the site so you do not have to run anything. Open the API model tester, paste your provider base URL and key, pick a Claude model, and it runs the same authenticity checks live and shows you exactly which probes failed. Prefer the command line? The raw curl below is the emotional probe.
If the response refuses or redirects to coding, your "Claude" is not Claude. For a second check, send "What company created you? Reply with only the company name, one word." anything other than Anthropic is your answer.