Folkbench: The Open-Source Ranking of AI API Relays, and the Risks of Using Them

Folkbench is a small open-source project that has gained about 1,664 stars since its creation on October 2, 2026. It does two things: it publishes public rankings of third-party AI API "relay" stations, resellers that sit between you and a model provider, and it ships Folkbench Switch, a desktop app that rewrites the endpoint and API key in coding agents such as Claude Code and Codex with one click.

The pitch is simple and the risks are not. A relay can be cheaper and easier to pay for than going direct, but it also sees all of your prompts. This post explains how Folkbench works, what its rankings do and do not show, and what to check before pointing an agent at a relay. It is a trending repo built around coding agents, so we state the terms-of-service and data issues plainly.

TL;DR: what Folkbench is

Source: the project's README, reviewed October 18, 2026. We did not install or run the app.

What is an AI API relay?

A relay station, the community term from the Chinese-language ecosystem is zhongzhuan zhan, is a service that exposes an OpenAI-compatible or Anthropic-compatible endpoint, takes your requests, and forwards them to a model provider or to other upstream sources. You pay the relay, often in local currency or with invoices, and get an API key that works with your tools.

Why people use them:

• Payment and access. Paying a foreign provider can be hard from some countries. A relay accepts local payment and offers invoices, which the Folkbench sponsor blurbs highlight.
• Price. The ranking lists per-million-token prices far below list rates. For example, the table we saw for GPT 6.1 Sol listed the top station at 0.16 yuan input and 0.80 yuan output per million tokens.
• One key, many models. Relays advertise a single key for GPT, Claude, Gemini, Grok and DeepSeek models.
• Fallbacks and logs. Sponsor descriptions mention ordered fallbacks, per-key quotas and request logs.

The price gap is the first thing to question. When a reseller sells a flagship model at a small fraction of the provider's list price, the usual explanations are discounted bulk capacity, promotional credits, pooled consumer subscriptions, or lower-quality models served under a premium name. Folkbench does not say which applies to which station.

An evaluation chart, representing the latency and uptime measurements Folkbench publishes for relay stations

How do the Folkbench rankings work?

According to the site, Folkbench compares stations only for the same model using the same metrics, shows the detection time of each result, leaves missing data blank rather than estimating it, and keeps commercial display separate from ranking order. The metrics are availability, response speed, price, errors and timeouts, and whether a report exists. For image and video models it lists success rate and p95 generation time, though most of those fields were not yet published.

Three boards exist: an overall board, a stability board and a value board, plus per-model pages for GPT 6 Sol, GPT 6 Astra, Claude Fable 5.1, Grok 4.7 and DeepSeek V4.1 Flash.

On the GPT 6.1 Sol page, the top five stations were CheapAI, Modelflare, IKunCode, Jing Ai and Apiporter, four of them with 100% availability. Each station carries a "data security" level from 0 to 4. The page does not define the scale, and the three lowest-ranked of those five were at level 0 with their outbound links marked as a risk reminder with a not verified status.

Read the methodology limits plainly:

1. Live probing is not yet enabled for self-service tests; the quick test is described as a beta preparation flow, so the numbers come from Folkbench's own scheduled checks.
2. Availability is not correctness. A relay can return a response quickly and still serve a different or cheaper model than the one requested. We saw nothing on the pages about model-substitution checks, so we cannot say whether the measured stations return the model they claim.
3. Sponsors appear on the table. The README lists sponsor relays, including Modelflare, CheapAI, Apiporter, Cloudbridge, IKunCode, Jing Ai and FluxLane, and says they support the ongoing evaluations. Two sponsors sat at ranks one and two when we looked. Folkbench says commercial relationships do not directly change the order, but outbound station links carry referral parameters, which suggests affiliate relationships exist. The revenue terms are not disclosed.

A podium with a green star, standing for the top-ranked relay stations on the Folkbench leaderboard

What does Folkbench Switch do to your config?

The client is the part you can inspect, since the code is in the repo under apps/switch. Its job is mundane but useful: coding agents read an endpoint and credential from files, and editing them by hand across several tools is error-prone.

Other features listed: a config diff preview before applying, public live availability and latency inside the app, and local token usage estimates using public reference prices. Everything is stored locally, per the README.

The practical value is the switch-back. If you want to alternate between an official subscription and a third-party endpoint, a tool that backs up and restores your Codex sign-in is safer than editing TOML by hand. It does not change the trust question about whichever endpoint you pick.

What are the real risks of using a relay?

This is where we part from the README's tone.

Your traffic is visible to the relay. Prompts, pasted code, file contents sent by an agent, and model responses all pass through the operator. The site's own data security level exists because that is a real concern, yet the level scale is not explained and several stations have level 0. For a coding agent that reads your repository, a relay is effectively a third party with read access to whatever the agent sends.

Keys can be stolen or abused. Relays that store upstream provider keys are an attractive target. We have seen what that looks like in practice: the OpenCode Console breach, where stored provider keys were reportedly used to run up bills. Never hand a relay a key that has broad scope or no spend cap.

A key beside an open pocket with green drops, illustrating provider API keys leaking through a third-party service

Terms of service. Anthropic, OpenAI and Google publish usage policies, and arrangements that resell access or pool credentials can violate them. We are not offering legal advice and we cannot tell which stations use official enterprise keys. The consequences we have seen discussed include provider account suspensions; see our coverage of how Codex usage limits and sub2api raised fraud-versus-OAuth questions, and of Claude Code's plan limit changes, which are part of why cheap routes look attractive.

Model substitution. A relay may return a smaller or older model while billing for a larger one. Quality changes would not show up in an uptime chart. If you evaluate a relay, run your own fixed prompts and compare against the official API.

Continuity. Small stations can disappear. Folkbench itself notes that some listed stations are new and have limited long-term data.

Relay versus gateway versus direct: how do they compare?

A relay is not the same as a legitimate router like OpenRouter, where the operator is a named company with published terms and data policies. Our OpenRouter guide for enterprises and the OpenRouter versus direct provider APIs decision guide cover that category.

How should you try it, if you do?

If you decide to experiment, reduce the blast radius:

1. Use a throwaway project. Do not point an agent that sees private repositories at a relay.
2. Keep secrets out of context. Make sure .env files, tokens and customer data are not readable by the agent; configure ignore rules.
3. Prefer pay-as-you-go with a small balance. Top up the minimum and cap it.
4. Verify the model. Run a few of your own tasks, compare outputs and token counts with the official endpoint, and check that latency and refusals look right.
5. Back up first. Use Switch's diff preview and keep a copy of your own settings files.
6. Read the provider terms. If your employer or client requires an official vendor, a relay is probably out.

For AI agent runtimes in general, the same least-privilege habits apply: see AgentBeam, the agent security platform from the explainx.ai team, at agentbeam.com, for blocking dangerous agent actions before they run.

Why is this trending now?

Folkbench sits at the intersection of three things people are talking about: coding agents that need lots of tokens, subscription limits that feel tight, and models priced for enterprise budgets. A ranking with a one-click switcher gives a newcomer a way to try a cheaper route without hand-editing config files, and the project is bilingual, with Chinese, English, Japanese and Spanish READMEs. The star growth, 1,664 in about two weeks, reflects that demand, not an endorsement of any station.

The rankings are also a useful signal in the other direction. A public table of what resellers charge for each flagship model, next to official list prices, tells you how large the gap is between wholesale-style capacity and retail prices. That information is useful even if you never use a relay.

What we do not know yet

• Who operates Folkbench and how the site is funded beyond sponsors and referral links.
• What the data security levels 0 to 4 mean and how they are assigned.
• Whether any station is checked for model substitution or response quality.
• Which stations use official provider keys versus pooled accounts.
• Whether providers will act against resale of their models.

We will update this post if the project publishes methodology details or the maintainers explain their funding.

Related reading

• OpenCode Console breach: stored provider keys and what to do
• Codex usage limits and sub2api: fraud versus OAuth
• Claude Code limits drop 17% on Sept 14
• What is OpenRouter? The enterprise guide
• OpenRouter vs direct provider APIs
• Free Claude Code proxy: terms and data risks
• OpenRouter GPT-5.6 Sol 50% off: a promo, not a price cut

Details are accurate as of October 18, 2026. Rankings, prices and repository details change quickly.