MODEL VERIFICATION

Do not take our word for it.

Use these five checks to compare a GlideAPI route with the official provider. Record the results and stop using any route that fails.

Get free credits to test Start the protocol

THE PROTOCOL

Five checks. One decision.

Run the same model, prompt, and parameters through both routes wherever the test calls for a comparison.

Verify model identity

For Claude reasoning models, preserve the returned thinking block and signature. Send that block back through Anthropic's official API as prior assistant context. An accepted signature is upstream-verifiable evidence.

You need

A GlideAPI key, an official Anthropic key, and the same Claude model available on both routes.

Pass condition

The official API accepts the signed thinking block without reporting an invalid or missing signature.

Send a reasoning request through GlideAPI and save the complete thinking block unchanged.
Place that block in prior assistant context in a request to the official Anthropic API.
Record the request IDs and the official API result.

{
  "type": "thinking",
  "thinking": "...",
  "signature": "UPSTREAM_SIGNATURE"
}

If it fails: retry once with an untouched block and matching model. A repeated signature rejection means the identity check did not pass.

Compare deterministic output

Send the same short factual prompt through both routes with matching parameters. Use a task with one verifiable answer rather than open-ended writing.

Keep identical

Model ID, system prompt, messages, temperature, seed when supported, token limit, and tool definitions.

Pass condition

The factual result and required structure agree. Wording may differ when the provider does not guarantee seeded determinism.

official = OpenAI(api_key=OFFICIAL_KEY)
glide = OpenAI(api_key=GLIDE_KEY, base_url="https://glideapi.dev/v1")

a = official.chat.completions.create(model=MODEL, messages=prompt, temperature=0)
b = glide.chat.completions.create(model=MODEL, messages=prompt, temperature=0)

If it fails: compare raw request bodies first. A different factual answer after parameters match is evidence to investigate, not proof by itself.

Compare token usage and billing

Prompt token counts should match for the same tokenizer and exact input. Compare returned usage, then verify the corresponding console line item.

Record

Input, cached input, output, reasoning tokens, model ID, request ID, and displayed charge.

Pass condition

Usage fields agree with the official route and the charge matches the live model rate within normal currency rounding.

Send one non-streaming request through both routes.
Save both usage objects before changing the prompt.
Calculate each charge from the published per-token rates and compare it with Usage in the console.

If it fails: confirm cached and reasoning tokens are not being compared with ordinary output tokens.

Check the full context window

Place a unique fact at the beginning of a near-limit prompt, then ask for it at the end. Run the same test against the official provider.

Test data

Use generated filler with a unique marker in the first block. Do not use sensitive files or repeated natural-language facts.

Pass condition

GlideAPI retrieves the marker at the same tested size as the official route and reports comparable input usage.

Start below the published limit and confirm both routes retrieve the marker.
Increase the prompt in measured increments without changing the marker.
Stop before the provider's documented hard limit and compare the last successful size.

If it fails: check client-side truncation first. Earlier failure only on GlideAPI means the context check did not pass.

Exercise every declared capability

Read the live catalog, then make a real request for every capability the selected model advertises. Do not infer support from the model family name.

Test separately

Image input, tool calls, JSON or structured output, streaming, and caching when each is declared.

Pass condition

Each advertised capability returns the expected response type, not merely an HTTP 200 response.

curl -H "Authorization: Bearer YOUR_GLIDE_KEY" \
  https://glideapi.dev/v1/models

Open the live model catalog

THE DECISION

If any check fails, do not trust the route.

This protocol exists to produce evidence, not another marketing claim.

Get free credits and run it Choose a model