Don't Trust Any Claude API Reseller.
Verify Them.

Including us. The reseller market is full of substituted models, silent downgrades, and cached responses billed at premium rates. Here are three checks — each takes minutes — that expose them. Run them against youragent, and against anyone else you're evaluating.

1. Check the model ID in every response

Every genuine Anthropic API response reports the model that actually served it:

curl https://your-agent.cc/api/v1/messages \
  -H "x-api-key: $YOUR_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-opus-4-8","max_tokens":64,
       "messages":[{"role":"user","content":"Say hi"}]}'

# Response includes:
# "model": "claude-opus-4-8"

If the model field is missing, rewritten, or doesn't match what you requested, walk away. This is necessary but not sufficient — a dishonest proxy can fake the field — so continue to check 2.

2. Run a side-by-side comparison

The hardest test to fake. Send the same prompt, same parameters, temperature: 0, to both endpoints:

• The provider you're testing (e.g. your-agent.cc/api)
• Anthropic's official api.anthropic.com (a $5 official account is enough)

Use a complex reasoning prompt — multi-step math, a subtle code review, a long-context recall task. Compare reasoning depth, formatting habits, and output length distribution. A substituted or downgraded model diverges visibly within a few hundred tokens.

3. Test behavioral fingerprints

Each Claude generation has behaviors that are expensive to imitate:

• Knowledge cutoff — ask about events only the claimed version should know.
• Tool-use formatting — newer models emit visibly different tool-call structures.
• Long-context recall — paste 50k+ tokens and query a detail from the middle; downgraded models miss it.
• Refusal style — each generation phrases boundaries differently.

What to do with the results

If a provider passes all three checks on the channel you intend to use, the remaining questions are operational: uptime history, support responsiveness, refund policy. If they fail even one — no discount justifies putting them in production.