Gemini API Relay Detection
Based on OpenAI Compatible Protocol + 7 dimensions, verifying Gemini API relay protocol compliance and capability completeness.
TokenAPI Scan is not a single-point speed test — it brings protocol behavior, usage fields, model responses, provider profiles, and leaderboard risk into the same evidence system.
https://hk.rootflowai.com/v1 — please confirm it matches your backend.https://generativelanguage.googleapis.com/v1beta/openai;
third-party relays typically expose the same protocol at /v1.OpenAI Compatible Protocol
Unified probing via POST /chat/completions, covering Google's official /v1beta/openai and most third-party relay /v1 endpoints.
Key Capabilities
Detects tool calls, JSON schema strict mode, streaming vs non-streaming response consistency — identifying missing capabilities or relay layers that don't pass parameters through.
Usage Fields
Checks usage.prompt_tokens / completion_tokens / total_tokens self-consistency and whether streaming include_usage is faithfully relayed.
Frequently Asked Questions
Why do Gemini API relays often return model_not_found?
Three possibilities: ① The relay only proxies a subset of Gemini models (many relays only have 3.x preview series, not 2.5); ② Model name spelling differences (gemini-2.5-flash vs models/gemini-2.5-flash); ③ The model has been deprecated. TokenAPI Scan runs a preflight before submission, identifying dead models within 500ms and suggesting available alternatives.
What should max_completion_tokens be when testing Gemini 3.x preview?
At least 64, recommended 128+. Gemini 3 enables thinking by default, consuming 30-60 reasoning_tokens before outputting text. If max is too small (< 32), thinking fills the budget and no space remains for text output — response becomes empty string with finish_reason=length. TokenAPI Scan's detector sets max to 64-384.
Why doesn't TokenAPI Scan support Gemini's native /v1beta/models/X:generateContent path?
99% of third-party Gemini relays only expose the OpenAI-compatible protocol, and Google also provides the OpenAI-compatible endpoint /v1beta/openai. Maintaining two separate detection logic sets is costly and prone to translation-layer distortion, so we focus on thorough coverage of the OpenAI-compatible path.
Can Google's official Gemini API also be tested with TokenAPI Scan?
Yes — fill in https://generativelanguage.googleapis.com/v1beta/openai with an AIza... key from Google AI Studio. Note that connecting directly to the official endpoint has network requirements; most visitors find relay access more stable. Check the Leaderboard for tested Gemini relays.
More Gemini relay questions → /faq#gemini · All FAQs
Popular Gemini / Google AI Provider Profiles
Sorted by tested model hits + pass rate:
Aihubmix
Hits 8 Gemini models · Pass rate 100.0%
gemini-3-pro-image · gemini-3.1-flash-image · gemini-3.1-flash-lite-image …
View Profile🔧 Network Environment Check
After verifying the relay's authenticity, local network variables still need to be ruled out. Below, we first read your browser's current public egress IP in-page, and use Net.Coffee + ProxyCheck for cleanliness reference — we do not mistake the server host IP for your egress.
AI Egress Cleanliness Reference
This shows public signals and third-party risk references, not equivalent to Claude / OpenAI official ban lists. For more granular network info (DNS, WebRTC, Ping, etc.), open the in-site toolbox.
Current Egress Profile
We first identify the public egress IP seen by the browser, then query third-party public signals and cache them to the local database.
Open IP ProfileNetwork Connectivity Check
Use in-site network tools to check latency, status, and browser access paths — reducing third-party redirects.
Open In-Site Tools