Content revised:
AI API gateway questions
Find practical answers about upstream access, billing, checks and privacy. These explanations help you assess options; they do not guarantee a provider's quality or the safety of deposits.
On this page
Basics and upstream services8 questions
What is an AI API gateway?
An API gateway connects you with an upstream model service. It forwards requests, provides access to multiple models, and unifies billing. It is useful because it is easy to connect, flexible on price, and can aggregate or distribute access. Using a gateway is not the same as calling a model's official API directly, so understand its real behavior and service risks before you choose one.
How does a gateway differ from an official API?
An official API gives you a direct service and billing relationship with the model provider. A gateway adds an operator with its own routing, billing and data-handling rules. The same model name does not imply the same contract, service assurances or privacy terms.
Are a token site and an API site different services?
These are informal labels, not reliable product categories. Check whether the offer is a callable API, a web account or subscription access, then verify models, billing units, limits and support terms. A balance labelled “tokens” may not represent the number of tokens actually processed by a model.
Why use an aggregated gateway?
Aggregation can reduce the work of managing multiple accounts, endpoints and bills, and make model comparison easier. A single endpoint does not automatically remove upstream failures or protocol differences. Validate the capabilities you need and keep an independently usable backup for important work.
When should I prefer an official service?
Prioritize official options when you need a clear data-processing contract, compliance commitments, direct support or an accountable service relationship. Official services also have quotas, regional restrictions and feature limits. Assess the actual contract and test your workload rather than assuming every requirement is covered.
How do official API and subscription-based upstreams differ?
An official API operates under its API contract and usage rules. Subscription-based access depends on a product's permissions, limits and terms; it is not interchangeable with a cash API balance. An operator's claim of an “official source” alone does not establish resale authorization, data protection or continuing capacity.
What do New API, Sub2API and CLI Proxy API describe?
These names commonly refer to different integration roles: New API for channel, key and billing management; Sub2API-style tools for adapting subscription access; and CLI Proxy API-style tools for adapting command-line client access into an interface. Roles vary by project, version and deployment. A name proves neither an upstream source nor permission to resell it.
Does a familiar panel or open-source program prove quality?
No. A panel is a management and presentation layer: the same software can connect different upstreams with different limits, operations and privacy rules. Open source does not prove that the deployed code matches the repository. Assess requests, bills, permissions and service history, not appearance alone.
Choosing a service and protecting funds6 questions
What risks do AI API gateways carry?
Common risks include responses that do not follow the official protocol, a swapped model identity, unusual knowledge behavior, rewritten responses, opaque logging or key handling, and availability that depends on a single upstream. Some sites also advertise a higher-tier model while forwarding requests to a weaker capability level.
If a check looks normal and a site ranks near the top, is topping up absolutely safe?
No. A check result and a leaderboard rank are not a guarantee of fund safety, and they do not guarantee that a provider will keep operating. Endpoint behavior, pricing, and service conditions can change.
Third-party gateways are strongly affected by upstream policy, and providers change frequently. Stay aware of the risk, start with the smallest top-up, and add small amounts only after the service has stayed stable. Avoid a large one-time top-up in case the provider shuts down unexpectedly.
Why can a low unit price still mean a high total cost?
Output usage, retries, caching, throttling, charges for failures and troubleshooting time all affect total cost. Compare actual debits and usable results on the same workload rather than a single multiplier. A detection check cannot audit every billing rule used by a provider.
Why do groups within one site behave differently?
Groups may use different upstreams, routes, limits, model configurations and prices. A result for one channel does not automatically cover every channel on that site. Record the model, group and time you actually use; do not combine one channel's low price with another's quality.
How should I arrange a trial and a backup?
Start with a restricted test key and the smallest necessary top-up. Run sanitized representative tasks and review errors, latency, usage and refund terms. Test the backup switch in practice, preferably with an independent upstream. A small trial limits exposure but does not remove risk.
What legitimate or risky factors can explain low prices?
Promotions, bulk purchasing, lower operating costs, different service tiers or permitted caching can lower prices. Other offers may involve unauthorized resale, stolen credentials, unexplained credits or model substitution. Price alone proves neither legitimacy nor wrongdoing. Verify authorization, terms and delivery, rather than seeking ways to obtain unauthorized access.
Running a check6 questions
How do I use the Hvoy AI check? Do I need to sign in?
The homepage check does not require registration or sign-in. Enter the API endpoint and a test-only API key, select the target model, then click “Start Detection”. Confirm that the endpoint, key, and model match one another before you start.
How can I check whether an AI API gateway is reliable?
Use Hvoy AI to send real requests to the endpoint, compare several aspects of its behavior for consistency, and review the reported warnings. A check is not official certification or a guarantee of long-term quality. Combine observations from different times with your own workload tests.
If you do not have token quota available, or you prefer not to run the check yourself, compare options on the Hvoy AI tested leaderboard.
How should the endpoint, key and model match?
They must belong to the same usable service, and the key must have access to the selected model. Follow the provider's documentation for the API base address rather than using a chat webpage URL. Confirm what a model alias maps to and never put the key in a URL or public text.
Can a check consume upstream credit?
A check makes real API requests, so your provider may deduct quota or charge for them. Usage depends on the model, enabled checks and provider rules; there is no fixed-cost promise. Set a suitable budget for the test key and review the bill afterward.
Does an error or unfinished check mean the model is fake?
No. An incorrect address, missing permission, insufficient credit, rate limits or network failure can interrupt a check. Review the configuration using the displayed guidance and retain sanitized errors and timestamps. Avoid blind repeated retries; a missing result is neither a pass nor proof of substitution.
How can I make a repeat check useful?
Keep the model, group and settings consistent, repeat a limited number of times at different moments, and record configuration changes. Separately test sanitized examples of your own tasks. This helps distinguish isolated issues from recurring behavior without treating a small sample as representative of every user or time.
Understanding results6 questions
How is the Hvoy AI comprehensive check different from a basic connectivity test?
A connectivity test mainly confirms that an endpoint can be called. The Hvoy AI comprehensive check looks beyond whether a call succeeds, and it does not rely on a single indicator. The report shows how the endpoint behaves and which risks need attention.
Can asking “who are you?” verify a model?
Not on its own. The answer can reflect system instructions, training material, context or provider rewriting, and a model may state its own version incorrectly. A claimed name is not official proof and should not replace multiple observations and workload testing.
Is a warning a final verdict of substitution?
No. A warning identifies an inconsistency worth reviewing in that observation. Version changes, settings, service limits or temporary failures may also matter. Read the complete report, check its date and repeat where appropriate rather than turning one warning into a final accusation against a provider.
Does one normal result guarantee future quality?
No. It describes the tested model and channel at that time. Upstreams, routes, limits and model versions can change. Continue observing your own success rate, costs and task quality after integration; a normal result guarantees neither ongoing service nor safe deposits.
Why look at both sample size and measurement time?
Small samples are sensitive to isolated events, while older data may no longer describe the current channel. Many samples do not cover every task or region, and a recent date does not imply enough evidence. Read the scope, sample size and actual probe time together; unknown values are not zero.
Models and clients5 questions
What should I check for Claude Code or Codex?
Start with the client's protocol, model scope, authentication and service terms, then review observations for the matching model and channel. Test tool use, streaming, long tasks and error recovery in the actual client. Passing an API check is not end-to-end client compatibility certification.
How should I shortlist services for Gemini?
Start with the exact target model and interface you use, not just a “supports Gemini” label. Confirm the required input types, context and tool capabilities, then test them with your client. A text-only observation cannot establish support for features that were not tested.
How should I compare Chinese models or a multi-model setup?
Verify the provider, exact version and billing specification, then compare quality, speed and actual costs on the same sanitized tasks. Do not merge different publishers or similarly named aliases. Suitability for Chinese text, code or long documents should be tested on your work, not inferred from a brand.
Does the same model alias guarantee the same capabilities?
No. An alias can map to different versions, configurations or routes, and mappings may change. Confirm the actual version and required capabilities. A name, self-reported response information or promotional screenshot cannot independently prove equivalence.
Does a check cover images, files, tools and every client?
Do not assume that. Each result covers only the checks actually run and their scope; a working text endpoint does not establish every other capability. List the features you require and test each in the intended client and channel. Untested functionality is not a pass.
Prices, multipliers and subscriptions9 questions
Why separate input and output prices?
Input and output can have different rates, and long answers, tool loops and conversation history change their proportions. Estimating a whole task from the lowest input price can be misleading. Account separately for actual input, output and any other applicable charges.
Which units must match before comparing quotes?
Align currencies and units such as per million tokens, per thousand tokens, per call or per duration, then check version, context tier and group. A missing quote is not free usage, and incompatible units cannot simply be added. A currency symbol on a panel may not identify the currency you actually pay.
Why can caching change the bill?
Cache writes, reads and misses may follow different pricing rules, while hit rates depend on the upstream, routing and request content. An ideal hit rate is not a guaranteed discount. Read the channel's cache policy and reconcile actual usage with balance debits.
Are platform currency and a group multiplier the same thing?
No. Platform currency describes internal balance units and their purchase cost; a group multiplier describes how requests consume those units. You need the real cost per unit, reference pricing and applicable multiplier to estimate cash spending. A “$” symbol or “0.1×” label alone does not establish a discount.
Why can the same 0.1Ă— multiplier produce different cash costs?
Illustrative arithmetic only: suppose a task with a 5 USD reference price is recorded as 5 platform reference units, with no other charges. At a purchase cost of 1 CNY per unit, 5 Ă— 1 Ă— 0.1 = 0.5 CNY; at 7 CNY per unit, 5 Ă— 7 Ă— 0.1 = 3.5 CNY. These are assumed conversions, not current exchange rates or offers. The 0.1Ă— label alone does not mean 90% off actual USD spending.
Does a monthly plan mean unlimited calls?
Do not assume so. Plans can limit frequency, concurrency, daily usage, models and permitted use, and may rely on shared resources or changing limits. Read expiry, overage, refund and authorization terms before paying, and include peak demand in your trial.
Is a subscription's “API-equivalent value” a cash balance?
No. A subscription buys access under defined terms, not cash that can freely be withdrawn, transferred or exchanged at official API rates. “Equivalent value” is usually an estimate based on assumed usage and reference prices. Limits, models, caching, utilization and terms affect it, and it does not automatically grant resale rights.
How should I read a subscription resale revenue example?
Hypothetical arithmetic only, not a real plan's yield or profit promise: assume a subscription costs 200 USD, notionally 1400 CNY. If a hypothetical 6400 reference units sell for 0.2 CNY each, revenue is 6400 Ă— 0.2 = 1280 CNY. That is already below 1400 CNY before other costs, including idle capacity, limits, payments, operations and refunds. It establishes neither real capacity, profitability nor resale permission.
How can I reconcile actual charges instead of trusting a multiplier?
Keep sanitized records for the same model, group and billing tier, then compare usage, input/output rates, cache records and balance changes. For discrepancies, first check units, retries and billing rules before sending key-free evidence to the provider. Hvoy AI model checks cannot detect or prove every billing anomaly.
Rankings and data definitions6 questions
Where can I see tested gateway rankings and hands-on reviews?
The Hvoy AI tested recommendation leaderboard shows measured results for listed sites, and the site directory shows hands-on usage notes. Use both together. You can also run a check on the homepage before connecting.
Why do monitored, displayed and priced site counts differ?
They describe different sets: observed sites, sites with public channels, recommendation-qualified sites, publicly priced sites and directory entries need not coincide. Counts deduplicate independent site records rather than merging equal names. No current public quote does not mean a site is absent from monitoring or the directory.
Do high uptime and low latency mean a service suits me?
Not necessarily. Uptime describes availability in the recorded samples; latency describes elapsed time for those test requests, not your region, client or long-running work. Uptime is not a model-authenticity percentage. Review samples and dates, then test your actual tasks.
Is a page update time the last test time?
No. The statistical window, actual probe time, quote observation time and page generation time have different meanings; a price update may also change a record's timestamp. Rankings include saved snapshots. Refreshing a page does not mean every channel was just tested, and unknown times should not be replaced with the current time.
How do advertising and organic measurements relate?
Treat advertising placement separately from measurements of organic channels. Buying advertising does not prove a passed check or recommendation eligibility. Sponsored display rows are excluded from organic statistics, while an advertiser's own organic channel may still qualify. This is not a promise that payment has no effect on any page placement.
Is a ranking an official endorsement or purchase guarantee?
No. A ranking organizes available observations under its display criteria; it is not model-provider endorsement, a complete audit or a refund promise. Combine dates, hands-on experience and a small personal trial. A ranking does not replace your contract with the service provider.
Privacy, safety and feedback6 questions
What should I check before submitting an API key?
Use a test-only key instead of an important production key, and limit that test key's quota or permissions. Do not expose the full key in public screenshots, comments, or shared content.
How should I manage a key after testing?
Where supported, give the test key minimal permissions, a small budget and an expiry. Revoke or rotate it when appropriate and inspect usage records. A normal check is not a reason to broaden access. If exposure is suspected, revoke it at the issuing provider; closing a page does not invalidate a key.
Should I test with production code, customer data or full prompts?
Avoid submitting them unchanged. Remove credentials, personal data, business secrets and unnecessary context; use representative but sanitized examples. Prompts, attachments and responses can all contain sensitive material. Being “just an API request” does not remove the need for a data review.
How should I assess a gateway's privacy risk?
Check retention, training use, access controls, deletion processes and contractual duties for both the gateway operator and upstream, and read the policies of the services you use. A check cannot prove that a provider keeps no logs or never uses data. Treat unverifiable claims as unverified and choose suitable contractual protection for sensitive work.
How can I report a questionable result or charge?
Use the site's contact page and provide the page address, model and group, time, reproduction steps and sanitized error or billing excerpts. Do not send full keys, personal payment credentials or unrelated private information. Also raise billing disputes with the provider that charged you; Hvoy AI cannot promise a refund on its behalf.