Unofficial and third-party LLM relays

Snapshot of pages opened on 14 November 2026. This note describes what those pages say a person might point an OpenAI-compatible agent at. A price, a “no KYC” line, or a “this is the real model” line is the site’s claim unless a second source on this page confirms it. Nothing here was checked by sending a prompt.

These are not the model makers’ own APIs. A base URL that speaks the OpenAI chat format only means a client can be pointed at it. It does not mean OpenAI, Anthropic, Google, or anyone else operates the server, stores the prompt the way their own policy describes, or will keep the route up.

1. gpt4free / g4f (xtekky)

Repository: xtekky/gpt4free. The README (fetched from the raw main branch the same day) says the project was created by @xtekky and is maintained by @hlohaus. It is licensed under the GNU General Public License v3.0. The README’s own summary is that the program is provided without any warranty.

The README describes GPT4Free as a community project that aggregates providers and ships a local GUI, a Python client, a browser client, and a FastAPI server that speaks an OpenAI-compatible inference API. The local example in the README is http://localhost:1337/v1 (slim Docker maps the API there). Model availability, the README says, depends on what each provider can do that day. The current list is deferred to g4f.dev/docs/providers-and-models.

What the docs say the public endpoints do

g4f.dev/docs/ready_to_use.html says the public collection is a set of base URLs for chat-style completions. Every chat call goes to {base_url}/chat/completions. The page’s table, as fetched on 14 November 2026, marks these as no API key required:

The same table marks https://g4f.space/v1 as key required: “Hosted instance, many models, get key from g4f.dev/api_key.” A later note on the same page says the hosted URL moved from host.g4f.dev/v1 to g4f.dev/v1, and that the /v1 endpoint requires a key from g4f.dev/api_key.html. Those two sentences do not name the same host. Treat the hosted URL as whatever that docs page says the next time it is opened.

The page also says a key is required for Azure, hosted instances, or gpt-oss-120b. Its client example still sends a placeholder api_key string even when the table says none is required. The same page lists ordinary vendor APIs (OpenAI, Google Gemini, xAI, Together, OpenRouter, DeepInfra, Nvidia’s own integrate host, and others) next to the g4f.space URLs. A URL on that list can be the vendor’s own API, which still needs that vendor’s key and follows that vendor’s bill.

https://g4f.space/api/auto is listed as “Auto Provider & Model Selection.”

Privacy of a public proxy

The hosted site’s privacy policy (version 1.2, last updated 22 October 2026) says g4f.dev is a non-commercial hobby project run by a private developer in Germany. The short version tells the reader not to send anything they would not share publicly.

When a chat message is sent, the policy says the prompt is forwarded to one of over 100 third-party endpoints. The chat UI is supposed to show which provider answered. If the user picks Auto, “the system picks a working provider automatically” and shows the one it used afterwards. The policy then says the operator has no influence on whether those providers log, store, or train on prompts, and: “Assume that every prompt can be stored by the receiving provider.”

The same policy says the operator may keep usage logs and excerpts of prompts and responses for up to 14 days for abuse prevention. Optional member-account data is kept at least 90 days. Browser chat history is described as staying on the device unless the user opts into an encrypted sync feature. That local-storage claim is about the g4f.dev chat page. It does not undo the forwarding: the prompt still leaves the device for whichever provider was chosen.

The README’s privacy section matches the local-storage claim and adds a production warning: lock down any server you run, and limit access to provider credentials and cookie/HAR storage. The README also says some adapters need API keys, and some need browser cookies or HAR files because those providers are reached through browser automation, plus Chrome or Chromium. A cookie or HAR file is a logged-in session. Pointing personal prompts at a public g4f.space URL means both the proxy operator and an unnamed upstream can read them. Running the library yourself does not remove the upstream.

Availability changes without notice

The hosted terms (also version 1.2, 22 October 2026) call the service an experimental, non-commercial aggregator offered “as is” and “as available.” They say there is no entitlement to availability, to a specific model, to a specific provider, to a speed, or to a quota, and that the operator may restrict, change, or discontinue the service at any time without notice. The operator says the service is not affiliated with or endorsed by the providers. The terms also forbid circumventing rate limits, bulk automated requests, and reselling access to the service or its providers.

The README already says availability depends on the provider. The January 2026 commit history on the repository is full of providers being removed. A base URL that answered yesterday can 404, or can answer from a different backend, with no changelog mailed to anyone using it.

2. NanoGPT

Pages opened: nano-gpt.com/pricing, the payment FAQ (bylined Milan de Reede, 14 October 2026), nano-gpt.com/provider, and the provider-selection docs.

Published billing

The pricing page offers two shelves.

Pay as you go. Listed as $0 per month. Top up a balance. The page says there is one free model, and that it is available on the web only, so it is not described as a free API tier.

$12 per month. The page calls this included usage on open-source models, on the web and through the API (it names SillyTavern and Janitor AI as examples). The published allowance:

• 60 million input tokens per week. Cached input counts in full.
• Output on the included models is free.
• 100 images per day, on 4 image models.
• Named open-source families: GLM, DeepSeek, Kimi, MiniMax, MiMo, Qwen, and more.
• A 5% discount on eligible paid text models. Paid models still bill the balance.
• Resets every Monday at 00:00 UTC.

The same box says the subscription does not include most closed models (it names Claude, GPT, Gemini, and Grok), video, voice, transcription, web search and other paid tools, some uncensored models (it names GLM 5.3 Uncensored), or choosing a specific provider.

Per-token claim. The page says text usage on the API and on the web is billed at list prices with no markup. Its worked example: GPT-5.5 costs $5 input and $30 output per million tokens at OpenAI, “and exactly $5/$30 here.” It says there is no percentage on top and no percentage fee on deposits, and that each chat-completion and Responses API response includes the exact cost charged. Two opt-in exceptions are printed next to that claim: manually pinning a specific provider adds 5%, and bring-your-own-key requests are billed 5% of the normal model cost.

The page also prints a “most used models” table in USD per million tokens (for example Gemini 4 Argon at $2 input / $10 output, Claude Opus 5.5 at $4 / $20, Claude Fable 5.2 at $8 / $40, DeepSeek V4.1 Flash 1017 at $0.10 / $0.60). Those rows are the page’s table on that day. The page itself says the full list loads with account pricing, so a logged-out snapshot is not the whole catalog.

Minimums and how you pay. The pricing banner says “start from $0.10 crypto or $1 card.” The payment FAQ repeats that: minimum $0.10 for every cryptocurrency, $1 for card. It says NanoGPT accepts 29 crypto options, including Monero, Bitcoin, Nano, USDC, and USDT, plus card through Stripe, including Apple Pay and Google Pay. It says every crypto deposit gets a 3% bonus (its example: $100 of USDT becomes $103 of balance). Balance is held in US dollars. It says a subscription is optional. It says crypto “needs no ID or account,” and that card payments are less private because Stripe processes the card and may run its own checks. That no-ID sentence is NanoGPT’s claim about its own checkout.

The platform picks the provider, and quantization can differ

nano-gpt.com/provider says: “NanoGPT auto-routes every request to a strong provider by default, and you can also pin a specific provider per request.” The provider list on that page spans many companies and regions. Pinning is the case the pricing page says adds 5% and excludes from the $12 allowance.

The provider-selection docs say defaultPrice is what you pay when you do not select a provider. They also say provider prices and availability can change between the moment you look them up and the moment you send the request.

Each provider row can include a quantization field, “weight precision served by this provider, such as fp8 or bf16.” The docs tell the reader to use quantization filters “when you need to avoid lower-precision provider routes for a model.” A min_quantization of fp8 still allows int8, fp8, fp16, bf16, and fp32. Integer and floating-point formats of the same bit width are treated as equivalent for that floor. So the default route can be a more compressed copy than another provider selling the same model name.

The same docs describe NanoGPT’s own check, for the most-used models that can be pinned to at least three providers: they send the same text to every provider and compare prompt token counts. Status failed means two or more checks in a row counted differently, “which usually means a different model.” That is their description of their own test. It is a token-count comparison, and they say it is omitted for models they do not check.

3. RelaySwitch

Pricing page opened: relayswitch.io/pricing. Also opened: relayswitch.io/no-kyc and relayswitch.io/proof.

Payment, KYC claim, model-check claim

The pricing page says every model is priced in USD per million tokens, input and output separately, with a lower cached-input rate where the model supports it. It says you prepay $10, $25, $50, $100, $250, or $500 by card or by crypto. The coins it names are USDT, USDC, BTC, ETH, and SOL. It says there is no subscription, credit does not expire, and there is no KYC. It points per-model prices at relayswitch.io/models and says that page covers all 349 models.

The no-KYC page says an email address opens the account (or Google or GitHub sign-in). It says a top-up from $10, in those same coins or by card, funds it. It says the account is not asked for identity documents, a phone number, a business registration, a developer account with the model makers, or a card on file for crypto top-ups. It says there is no subscription and billing is per token. It says request and response bodies are not stored, and that the account is identified by its email address.

Those are RelaySwitch’s claims about its own signup and logs. The upstream that runs a given model still receives the prompt; the page does not claim otherwise.

The pages do not agree with each other about the catalog size. On the same fetch, the no-KYC page’s opening sentence says 349 models, a summary box says 322 models on one key, and the base-URL line says OpenAI, Anthropic, and Gemini-format access on 313 models. The proof page header, generated 2026-11-14T13:26:44.569Z, says 358 models checked. Use the live catalog, not any one of these sentences.

The no-KYC page prints a “flagship” table it says was read from the catalog at render, with its own “under list” percentages. Examples it showed that day, in USD per million tokens: Claude Fable 5.2 at $4 input / $20 output against a list it states as $8 / $40; Claude Opus 5.5 at $2.716 / $9.70 against $4 / $20; GPT-6 Luna at $0.0625 / $0.20 against $0.10 / $0.50. One-message examples on the same page: $0.00031 on Manta Flash 1.0, and $0.0033 on Claude Sonnet 5.5, for a turn the page describes as 2K in / 300 out. These are the site’s comparisons. They were not checked against the model makers’ price pages for this note.

What the “real model” check claims

relayswitch.io/proof says every model gets the same canary prompt once a day, and the answer is compared with earlier runs. “A relay that swaps the model behind an id shows up here as a change.” Verdicts, in the page’s words:

• Stable. The model answered the canary the way it did on the earlier runs.
• Changed. The answer moved. “What serves the id is not what served it before.”
• Not reached. No route answered the latest run. The last verdict is history.

The snapshot header that day: 358 checked, 325 stable, 20 changed, 13 not reached. The page also says each request carries a signed receipt (model, billed tokens, charge) and that the same proof data is public JSON at https://api.relayswitch.io/v1/proof.

Read that check as what RelaySwitch says it does. A stable row means the canary answer matched earlier runs of RelaySwitch’s own prompt. It is not an independent weigh-in of the model file. Twenty models were already marked changed on the day this page was opened.

4. Chutes

Pages opened: chutes.ai/pricing and chutes.ai/privacy.

Free tier, how you pay, markup

The pricing page headline says: “No middlemen. No markup tiers. Pay for the tokens you use, or deploy your own chute on a confidential GPU and pay by the second.”

Under model pricing it says: “Pay per token. No subscription, no minimum, no markup.” The FAQ says billing is pay-as-you-go: top up a balance and pay per-token rates for hosted models, or per-second GPU time for a private chute, “with no subscription and no minimum commitment.” It says per-model rates are pulled live from their pricing API, “so what you see is what you pay.”

On the free tier, the FAQ is direct: “We do not offer a free tier at this time. Top up your account in the app to get started.” It also says monthly plans are no longer sold on chutes.ai. An existing plan, if someone still has one, runs to the end of the current period and does not renew; after that, usage is pay-as-you-go.

Payment, from that FAQ: “25+ payment methods via Stripe including crypto, cards, and bank transfers.” Also TAO and Bittensor alpha tokens. Enterprise invoicing is offered as a separate conversation. The “no markup” sentence is Chutes’ description of its own rates. The page still sets a price per model; it is not claiming those rates equal each model maker’s list.

The table on the page is live and long. One row it showed that day: google/gemma-4-31B-turbo-TEE at $0.12 input and $0.37 output per million tokens, cached input $0.012. Another: Qwen/Qwen3.8-27B-TEE at $0.039 input and $1.25 output. Private GPUs are a different product: the page showed an RTX Pro 6000 class at $2.25 per hour plus a one-time deployment fee of 3× the hourly rate ($6.75 in that example). Those figures are the page’s figures that day.

Who the privacy page says can read the prompt

The privacy page says Chutes does not log, store, or persist API request or response content on Chutes systems, subject to exceptions it lists (transient buffers, content the user saves in Chutes’ own apps, legal holds, content the user asks them to debug). It says they do not train on inputs or outputs. It also says API servers may read request content in memory to route it, unless the user is in an end-to-end encryption mode.

It splits models in two:

• Confidential compute / TEE. The catalog flag confidential_compute: true. The page says these run on independent operators on the Chutes subnet of Bittensor, inside hardware enclaves, and that the design is meant to keep the machine’s operator from reading the content. End-to-end encryption is an extra mode on top of that.
• Standard models. Not in a Chutes TEE. The page says many of these are served by third-party inference platforms. Chutes sends the inputs to that platform. “A Model Provider’s own staff and systems can access content while they process it.” Retention, as the page describes the providers’ commitments, runs from zero-retention (cache generally gone within 24 hours) to standard retention “generally up to 30 days,” and longer for safety flags. The page says those are contractual commitments Chutes cannot guarantee, because it does not operate those systems.

It also says an automatic or fallback route may select a standard model, and that a confidential model has to be requested by its identifier if every request needs the TEE claim.

5. Discount relay stations (“API relay”, 中转)

This is a category, not a recommendation. Shops in China and elsewhere sell a prepaid API key, tell the user to change the OpenAI base URL, and advertise a price below the model maker’s list. The usual shape:

• Prepaid key. You buy a balance. The unit is often a “dollar of credit,” which may not be a dollar of cash. A site can charge several yuan per dollar of credit, then add a bonus at higher top-ups. The number on the model table is not the cash you paid until that exchange rate and bonus are included.
• Prices below list. The shop publishes a multiplier (“0.2× official”) or a per-model rate under the vendor’s posted price. Groups and channels on the same shop can carry different multipliers for the same model name.
• Unknown logging. The prompt, the files, and the key pass through the shop’s server before they reach whoever actually runs the model. Some shops publish a privacy policy. Many do not say how long prompts are kept, who the upstream is, or which country the machine is in.
• No SLA. Channels are added and pulled as upstream accounts get rate-limited, shut, or repriced. The terms that exist usually say a model may disappear and that the shop is not the model maker.

apiprice maintains a public directory of these shops. Its methodology page, updated 14 November 2026, says it only writes down what it can re-read on a public page, and: “收录不代表推荐，也不代表安全” (being listed is not a recommendation, and it is not a statement that the shop is safe). On price it says the multiplier is the sticker divided by the vendor list price, and then: “倍率只反映标价，不反映你实际拿到的是不是这个模型、请求会不会失败。” The sticker does not show whether the model you received is the model you named, or whether the request will succeed. The same page says apiprice has not tested model identity, uptime, refunds, or how the shop handles data. It also says resale can conflict with the vendors’ own terms: it points at OpenAI’s rules against reselling or transferring API access and against use outside supported countries (it says mainland China and Hong Kong are not on OpenAI’s supported-country list), Anthropic’s rule against resale without approval, and Google’s Gemini API rule that the API client is offered in available regions. A cheap route can die because the upstream account was closed, and the person who bought the credit has no contract with the model maker.

A named example with a public price page: OhMyGPT

OhMyGPT (ohmygpt.com, operator named on its own terms as DogeNet LLC) is a long-running shop in this category. Its homepage, cost docs, privacy policy, and terms were open on 14 November 2026. Listing it here documents those pages. It is not an endorsement, and the same caveats apply to any similar shop.

The homepage says one key reaches 294 models and that, if you do not name a supplier, OhMyGPT picks one and fails over to another if the first errors or is rate-limited. That is the shop choosing the route.

Price caveat, in the site’s own words and numbers. The homepage section “比官方价便宜” says 96 models are below the official price, and that the comparison is a blended price per million tokens with cache hits, input, and output weighted 7:2:1. A “cheaper than official” badge on that page is that blend, not a promise that every input token and every output token is the vendor’s list price times one fraction. Some rows put the discount in the product name itself (GPT-6.1 Sol-10%, DeepSeek V4.1 Flash 1017-55%). A Gemini 4 Hydrogen row that day showed $1.35 input and $6.75 output per million, against a “标价” of $1.50 and $7.50, labeled −10%.

The cost docs add a second caveat. Billing is in credits, and “1 美元 = 250,000 Credits.” Each line is rounded up to a whole credit. Their own worked example says the unit prices on a tiny request were 45% of list, while the request itself cost 3 credits against a list of 4. They say a normal-sized request lines up with the advertised ratio. Small calls, and any call where the published ratio is a blend, can cost a different share of list than the homepage badge.

The privacy policy (last updated 11 September 2025) says OhMyGPT reserves the right to store, review, and investigate communication data in the system, that an upstream such as OpenAI may also record and review it, and that OhMyGPT usually cannot stop the upstream from doing so. It says they will not sell interaction inputs and outputs. The terms (last updated 22 October 2023) say the models are supplied by third parties such as OpenAI and that the company does not produce its own proprietary models; that the user accepts a no-refund policy except where the company cannot continue service because of force majeure; and that the user grants DogeNet LLC a permanent, worldwide, sublicensable copyright license over text and image prompts entered into the service and over assets the service generates from those prompts. The terms also say DogeNet LLC is not liable for direct or indirect loss from the service.

6. Risk table

“Apparent cost” is what the opened pages advertise. It is not a measured bill.

Sources

Opened 14 November 2026.

• gpt4free README: https://github.com/xtekky/gpt4free and https://raw.githubusercontent.com/xtekky/gpt4free/refs/heads/main/README.md
• g4f ready-to-use endpoints: https://g4f.dev/docs/ready_to_use.html
• g4f privacy policy (v1.2, 22 October 2026): https://g4f.dev/privacy.html
• g4f terms (v1.2, 22 October 2026): https://g4f.dev/terms.html
• NanoGPT pricing: https://nano-gpt.com/pricing
• NanoGPT payment FAQ (14 October 2026): https://nano-gpt.com/faq/payment-methods-and-cryptocurrencies
• NanoGPT providers: https://nano-gpt.com/provider
• NanoGPT provider-selection docs: https://docs.nano-gpt.com/api-reference/miscellaneous/provider-selection
• RelaySwitch pricing: https://relayswitch.io/pricing
• RelaySwitch no-KYC page: https://relayswitch.io/no-kyc
• RelaySwitch daily check: https://relayswitch.io/proof
• Chutes pricing: https://chutes.ai/pricing
• Chutes privacy: https://chutes.ai/privacy
• OhMyGPT homepage: https://www.ohmygpt.com/zh-CN
• OhMyGPT costs: https://www.ohmygpt.com/docs/zh-CN/costs
• OhMyGPT privacy (11 September 2025): https://www.ohmygpt.com/zh-CN/privacy
• OhMyGPT terms (22 October 2023): https://www.ohmygpt.com/zh-CN/tos
• apiprice relay methodology (updated 14 November 2026): https://apiprice.app/zh/relays/methodology/

A plain warning

A cheap base URL is someone else’s server sitting between you and the model. That server can read the prompt, keep it, and answer from a smaller or more compressed model while the bill still shows the name you asked for. The route can vanish overnight, the prepaid balance can vanish with it, and the company that actually made the model can shut the upstream account because resale breaks its rules. Keep passwords, private code, customer records, and anything you would not paste into a public chat off these relays. If the answer matters, or the data is yours to protect, use the model maker’s own API and pay the price on their site.