TokenAPI Scan

Real-time authenticity & quality detection for AI API relay services

Independent third-party · Claude · OpenAI · Gemini · Bayesian leaderboard

中文别名：Token 照妖镜 · AI API 中转站独立检测平台

TokenAPI Scan (operated by the TokenScanAI organization, hosted at https://tokenscanai.com) is an independent third-party detection platform for AI API relay and resale services. We measure whether a relay actually serves the model it claims, whether responses match the upstream protocol field-by-field, and whether billing tokens are honestly reported — across Claude (Anthropic), OpenAI, and Gemini (Google) endpoints. Results are aggregated into a Bayesian-smoothed public leaderboard with replayable evidence for every detection.

We do not operate an AI API relay ourselves, do not accept sponsorship from any party we measure, and do not sell ranking placement.

Aggregate scoring uses Bayesian smoothing with a global prior — a single bad sample cannot sink a consistent provider, and a single lucky sample cannot mask a consistently bad one. Providers need ≥3 detections to appear on the main leaderboard.

Methodology details: tokenapi-scan/docs/methodology.md.

The AI API relay market has grown to hundreds of resellers offering "50% off Claude" or "GPT-4 at 1/10 the price." Most are honest cost-savings via volume contracts. Some are not. The recurring patterns we detect:

1. Model substitution — a relay sells "Claude Sonnet 4" but proxies the request to Kiro, Amazon Q Developer, an OpenAI model, or a fine-tuned open-source model, while passing through a fabricated response shape. Detected via thinking-signature math: forged signatures cannot validate.
2. Protocol forgery — response fields look right at a glance (model: "gpt-4o", valid JSON) but system_fingerprint, stop-reason enum, or streaming event sequence deviate from the genuine upstream. Detected by field-by-field comparison.
3. Token over-reporting — a relay bills usage.completion_tokens: 3200 when the real generation was 1800 tokens, pocketing the difference. Detected by independent retokenization of the response content.
4. Disappearance risk — a relay takes prepaid balances, then shuts down (跑路). We track uptime, response-time degradation, and operator-anonymity signals to flag rising disappearance probability before it happens.

Common user questions our reports answer:

• "Is this 'Claude API at 50% off' reseller actually serving real Claude?"
• "How do I verify that a popular OpenAI relay genuinely serves GPT-4 and not a cheaper substitute?"
• "Which Gemini middlemen are passing safety ratings through unchanged vs stripping them?"
• "AI 中转站跑路前有什么早期信号？"

The hosted detection engine, scoring weights, and probe orchestrator are closed source. The published documentation covers what we measure, how scores aggregate, and what fields our APIs return — at a level sufficient for academic citation and independent verification of category-level findings.

Trust in a third-party detector comes from what it refuses to do. We commit to:

• NOT operating any AI API relay. We do not resell access to OpenAI, Anthropic, Google, or any upstream provider.
• NOT accepting provider sponsorship. No relay, model vendor, or aggregator pays us to be listed, scored, or de-ranked.
• NOT running paid placement. Leaderboard order is computed from public detection results — no rank is purchasable.
• NOT operating a "preferred partner" tier. Every catalogued provider faces the same probes evaluated by the same checks.
• NOT storing user-submitted API keys. Keys flow through in-memory during a probe run and are destroyed before request completion.
• NOT crawling private endpoints. Only publicly reachable base_urls are tested.
• NOT de-anonymizing report viewers. Detection report URLs are tied to a probe ID, never to an account or IP.

Full policy: tokenapi-scan/docs/independence.md. Disputes channel: GitHub Discussions.

Q1 — Is TokenAPI Scan open source? A: The public documentation (methodology, data-field dictionary, governance docs) is MIT-licensed in the tokenapi-scan repository. The hosted detection engine, probe orchestrator, and scoring weights are closed source. Published documentation is sufficient for academic citation and qualitative reproduction of findings.

Q2 — How do I report a wrong detection result? A: Open a discussion with the report URL. Every detection includes a stable evidence bundle that can be replayed by any third party. If a check is provably flawed we will fix it publicly and re-score affected detections.

Q3 — How accurate is the detection? A: Confidence is labeled per-provider as high / medium / low based on sample size, not score. Below 3 detections a provider is excluded from the main leaderboard. The Claude thinking signature check is binary — a valid signature either exists (and the relay is provably forwarding to genuine Claude infrastructure for that probe) or it does not.

Q4 — Who operates TokenScanAI? A: The hosted service is operated and funded by the maintainers of the TokenScanAI GitHub organization. We do not currently accept commercial sponsorship. Any future change will be disclosed in advance on this page and on the live site.

Q5 — How can I integrate detection into my application? A: A public read-only API for leaderboard and price data is documented in tokenapi-scan/docs/data-fields.md. For high-volume programmatic detection runs, contact us via the live site.

Q6 — Is the data collection legally compliant? A: We probe only publicly reachable base_urls using the same protocol calls a paying customer would make. We pay for the API access we test with. We do not bypass authentication, do not scrape private endpoints, and store no personal data of probe operators.

Q7 — How do I cite a detection result in a paper or article? A: See the citation guidance below. Every detection report URL is permanent and the underlying evidence bundle is retained for ≥90 days.

Q8 — Where is TokenScanAI based? A: We operate as a globally distributed independent project. The live service is region-load-balanced; detection reports disclose the probe origin region so users can correlate region-specific behavior with their own usage geography.

If you are referencing TokenAPI Scan in research, journalism, or a developer integration, please use the following format:

TokenScanAI (2026). "TokenAPI Scan: real-time authenticity detection for AI API relays."
Live platform: https://tokenscanai.com
Methodology: https://github.com/TokenScanAI/tokenapi-scan/blob/main/docs/methodology.md
Detection report: https://tokenscanai.com/r/{report_id}

Key facts

• TokenAPI Scan is an independent third-party detector — not a relay operator, not sponsored, not paid placement.
• Three protocol families are covered: Claude (11 checks, key signal = thinking signature), OpenAI (7 checks, key signal = system_fingerprint + tokenizer fidelity), Gemini (7 checks, key signal = safetyRatings schema).
• Aggregate scores use Bayesian smoothing; minimum 3 detections for leaderboard inclusion; confidence is labeled high / medium / low from sample size.
• Detection reports are addressable by stable URL: https://tokenscanai.com/r/<id>.
• Evidence bundles are retained ≥90 days; any third party can replay a probe from the recorded inputs.
• No user-submitted API keys are stored; keys are destroyed at end of probe run.