Model A, anonymousstreaming
Route by policy, not by one metric. Cheap models handle low-risk utility calls; verified channels handle user-facing or regulated output; latency caps protect interactive flows.
AI API Trust Layer, not another cheap proxy
Compare, benchmark, and verify AI APIs before production.
youtoken.cc helps developers test any OpenAI-compatible endpoint with their own key, compare model output in a blind arena, and publish reproducible reports. We do not store BYOK secrets. We do not call verification absolute truth. We show evidence.
BYOK
Benchmark without saving provider API keys.
Elo
Blind Arena voting updates transparent rankings.
Risk score
Verifier reports probability, evidence, and limits.
Reports
Shareable snapshots include time, node, model, params.
Model B, anonymouscomplete
Cost is only useful after confidence is known. The system should record provider behavior, compare error modes, then pick the lowest-cost route that clears the trust threshold.
#1Claude Sonnet 4.51548+18
#2GPT-4.1 mini1514+7
#3DeepSeek V3.21496-4
#4Gemini 2.5 Flash1479+3
Three trust tools first. API resale later.
The MVP earns confidence before charging for routing. Each tool produces evidence users can inspect, share, and challenge.
Blind model comparison
Two models answer the same prompt anonymously. Users vote on usefulness, then the system reveals names and updates Elo.
Latest match
gpt-4.1-mini beat deepseek-chat by 61% preference over 93 votes.
BYOK endpoint benchmark
Paste a Base URL and temporary key for the current browser session. Reports include TTFB, tokens/sec, total latency, node, and params.
TTFB 318ms / 46 tok/s
Provider claim verification
Run knowledge cutoff checks, response pattern analysis, optional logprob fingerprinting, and output a probability score with caveats.
Confidence 74%, medium risk
Public reports are the growth loop.
Every useful test becomes a shareable artifact with enough metadata to reproduce or dispute the result.
perf/kr-seoul/gpt-4.1-mini/2026-05-21
Node: Korea VPS, cold start: no, concurrency: 1, prompt hash: 8f4c1a
public
verify/provider-alpha/gpt-4-claim
Evidence: model list, cutoff questions, style fingerprint, user appeal open
medium risk
arena/code-generation/week-21
2,418 votes, session/IP limits enabled, Elo K=32, new model K=64
ranked
Report Service
Reproducible by default
Reports include model, endpoint class, timestamp, geographic node, parameters, tool version, limitations, and appeal status. No raw API keys. No private prompts in public reports unless the user explicitly publishes them.
Commercial API routing stays behind the compliance gate: legal entity, payment path, privacy terms, abuse limits, and upstream cost model must exist first.
Trust products must explain themselves.
The methodology page is part of the product, not documentation nobody reads. It tells users where the data is strong and where it is not.
Status, risk, and operations are first-class.
Developers will not trust a trust platform that hides incidents. The status board becomes the public operating ledger.
Implementation path
The static page ships the story now. The product grows only where trust and compliance are ready.
Static trust prototype
Replace the old identity positioning with AI Trust Layer landing, tools mock, methodology, status, and waitlist.
Working tools
Build Arena, BYOK Perf, Verifier, public reports, limits, and provider appeals without payments.
Growth back office
Launch tools directory, code snippets, TrustMark workflow, audit log, SEO report pages, and moderation.
Commercial API
Only after entity, payment, compliance, cost model, abuse ceilings, and observability are verified.
Join the trust layer before it becomes a gateway.
Early users get Arena voting, BYOK benchmarking, verifier reports, and provider retest workflows first. Paid routing comes later, after the boring parts are real: entity, tax, privacy, audit logs, risk controls, payment rails.