Research

Gray-market OAuth and relays

Checked 14 November 2026. This chapter combines the gray-market OAuth note, the gray-market relays note, and the Anthropic and Gemini sections of the OpenClaw note. It describes published terms, enforcement reports, and what the relay sites claim. A price, a “no signup check” line, or a “this is the real model” line is that site’s claim unless a second source here confirms it. Nothing here was checked by sending a prompt.

A subscription login is the “sign in with Google / Claude / ChatGPT / GitHub” session. An API key is a separate credential that bills a developer account. A relay is someone else’s server that speaks a familiar chat format and sits between the program and whoever actually runs the model. This chapter is not a setup guide.

Google Gemini CLI and Antigravity

Google’s Gemini CLI terms, last updated 10 April 2026, name the practice:

Directly accessing the services powering Gemini CLI (for example, the Gemini Code Assist service) using third-party software, tools, or services (for example, using OpenClaw with Gemini CLI OAuth) is a violation of applicable terms and policies. Such actions may be grounds for suspension or termination of your account.

Source: Gemini CLI: License, Terms of Service, and Privacy Notices. The same site’s FAQ, also last updated 10 April 2026, says using third-party software “to harvest or piggyback on Gemini CLI’s OAuth authentication” is a direct violation and “may be grounds for immediate suspension or termination.” The method it names for a third-party coding agent is a Vertex AI or Google AI Studio API key. Source: Gemini CLI FAQ.

Antigravity, the consumer product Google later pointed those users toward, has its own sentence. Section 6 of the Google Antigravity Additional Terms of Service says:

Using third party software, tools, or services to access the Service (e.g. using OpenClaw with Antigravity OAuth) is a breach of this Agreement. Such actions may be grounds for suspension or termination of your Antigravity and/or Gemini CLI accounts.

The Antigravity FAQ repeats that Claude Code, OpenClaw, or OpenCode with an Antigravity login violates the terms and can suspend or terminate the account, and it recommends a Gemini Enterprise or Google AI Studio API key. Wrapping the Antigravity CLI inside a third-party agent, or automating account rotation, “is not supported under consumer plans.”

February 2026 suspensions

On 27 February 2026, Gemini CLI developer relations posted in the official Gemini CLI repository that account disruptions that week came from “a series of Antigravity bans” for “the use of 3rd party tools or proxies to access Antigravity resources and quotas.” Because abuse prevention sat on a shared backend, those bans also blocked Gemini CLI and Gemini Code Assist. Google said it was unbanning recently flagged accounts and moving to a recertification form. A second terms violation “will be permanently banned.” The same post restates that piggybacking on Gemini CLI OAuth is a direct violation. Source: Addressing Antigravity Bans & Reinstating Access.

The Register’s 6 April 2026 account of the same February action attributes that piggyback sentence to Jack Wotherspoon, Gemini CLI developer relations. Source: Anthropic closes door on subscription use of OpenClaw.

Google’s deprecation page, last updated 9 October 2026, says that starting 18 June 2026, Gemini Code Assist IDE extensions stopped serving the individuals, Google AI Pro, and Google AI Ultra tiers, that this also applies to Gemini CLI, and that “you can no longer use the Login with Google option to access the IDE extensions or Gemini CLI.” Consumer users “can migrate to the Antigravity family of products.” Standard and Enterprise are called out there as a separate schedule. Source: Gemini Code Assist consumer accounts.

The April Gemini CLI terms page still lists “Logging in with your Google account to Gemini Code Assist” as a supported method. The October deprecation page says that consumer Login with Google path stopped being served on 18 June 2026. This chapter treats the later deprecation page as the one that names the shutdown date. It does not treat the April terms page as proof that a personal Google login still opens Gemini CLI.

OpenClaw’s current Google provider page says:

OpenClaw does not offer new Gemini CLI OAuth or Antigravity OAuth setup. Google ended consumer Gemini CLI Login with Google access on June 18, 2026, and the Antigravity terms prohibit third-party tools from accessing the service through Antigravity OAuth. Use an AI Studio API key or Vertex AI instead.

The same page says existing valid Gemini CLI OAuth profiles “remain executable for compatibility, but OpenClaw cannot create or repair them,” and that a broken one is to be replaced with an AI Studio API-key profile. An optional local Gemini CLI runtime stays on the AI Studio API-key path. Source: OpenClaw Google (Gemini), also recorded in the OpenClaw note. This chapter does not describe how to keep an old login alive.

Anthropic

Two Anthropic statements both appear in the source notes. This section keeps both. They are not the same sentence.

On 15 June 2026 Anthropic paused the previously announced Claude Agent SDK change. The help page Use the Claude Agent SDK with your Claude plan, quoted in the OpenClaw note and summarized in the OAuth note, says:

Update June 15, 2026: We've paused the previously-announced changes to Claude Agent SDK usage. For now, nothing has changed: Claude Agent SDK, claude -p, and third-party app usage still draw from your subscription limits.

The same page’s 7 October 2026 update says Max, Ultra, and Team plans include monthly API credits for API-key use, and that “when you’re signed in with your Claude plan, the Claude Agent SDK and claude -p still draw from your subscription limits.” That October sentence names the Agent SDK and claude -p. It leaves out the words “third-party app usage.”

The current statement about third-party tools on a subscription login is the help page Log in to your Claude account, updated in the week before the OAuth note was checked. It still says Anthropic may draw third-party tool use from usage credits, and that the user pays those charges:

The preferred way to access Anthropic services using third-party software, tools, or services (“third-party tools”), including open-source projects, is through API key authentication through Claude Console or a supported cloud provider. Anthropic may at its discretion allow paid subscribers who have enabled usage credits to use certain third-party tools to access Anthropic services included in paid subscription plans, but reserves the right to draw use of such third-party tools from usage credits rather than subscription limits. Users are responsible for any usage credit charges incurred this way. Use of third-party tools that misrepresent their identity to Anthropic’s servers, attempt to route third-party traffic against subscription limits, or otherwise violate applicable terms or policies is prohibited and such use may be enforced against.

The Usage Policy, effective 12 November 2026, tells users not to “Resell, proxy, or otherwise provide access to Claude through unauthorized means, including services that route requests through consumer subscriptions or misrepresent the product or client being used.” Source: Usage Policy (PDF). Claude Code’s legal page says OAuth “is intended exclusively for purchasers of Claude Free, Pro, Max, Ultra, Team, and Enterprise subscription plans and is designed to support ordinary use of Claude Code and other native Anthropic applications,” and that Anthropic does not permit third-party developers to offer Claude.ai login in their own applications or to route requests through Free, Pro, Max, or Ultra credentials on behalf of their users. Source: Claude Code legal and compliance.

The earlier announcement is the context for the June pause, and it is a different date. The Register, 6 April 2026, quotes Boris Cherny, head of Claude Code: “Starting tomorrow at 12pm PT, Claude subscriptions will no longer cover usage on third-party tools like OpenClaw.” An Anthropic spokesperson said that starting 4 April, third-party tools would draw from extra usage instead of subscription limits, that using Claude subscriptions with third-party tools is not permitted under the Terms of Service, and that customers could still use Claude with third-party tools through extra-usage bundles or an API key. The same article says Claude subscriptions continued to apply to Claude.ai, Claude Code, and Cowork. Source: The Register, 6 April 2026. The OpenClaw note records that those April news accounts are not the current help-center text. The 15 June 2026 page is the later pause.

Extra usage remains a billing switch in the OpenClaw note. Buy usage bundles says bundle balance “works across Claude, Claude Code, Claude Cowork, and third-party products” and “only apply after you've exceeded your plan's limits and are using usage credits.” Prices on that page when it was checked: a $50 bundle costs $45, a $250 bundle costs $200, a $1000 bundle costs $700. Monthly discounted purchase caps: $2000 for Pro, Max, and Ultra; $3000 for Team owners. Past the cap, usage is “billed at standard rates.” The page says prices can change. OpenClaw’s usage-tracking page still shows an optional extra-usage budget for subscription credentials. If usage credits are turned on, going past the plan can cost money. A separate OpenClaw FAQ line, Extra usage is required for long context requests, means the request wants Anthropic’s 1M context window and the credential is not eligible for long-context billing.

OpenClaw’s OAuth page says Anthropic staff told the project that Claude CLI reuse is allowed again, and a costs page says that as of 5 July 2026 Anthropic had confirmed Claude CLI reuse, including claude -p, as a sanctioned pattern unless it publishes a new policy. OpenClaw also says Anthropic can change the June pause without an OpenClaw release, and it still calls an API key the safer production path. That staff assurance was not on the Anthropic pages fetched for the OAuth note. A reader who wants to know the bill has both of the statements above: the 15 June 2026 pause, which says third-party app usage still draws from subscription limits, and the login page, which says Anthropic may draw that use from usage credits and the user pays those charges.

The pages checked support a conditional allow for OpenAI’s published sign-in, and they forbid turning that login into a shared or general-purpose API. Nothing fetched for the OAuth note was an OpenAI page that names OpenClaw.

Sign in with ChatGPT Terms, effective 4 November 2026, govern an application that lets a user connect a ChatGPT account and, where enabled, use that plan in the app. Tokens must be obtained through OpenAI’s supported sign-in flow. Persistent storage “must be local and under the user’s control, not in a remote or managed environment.” If the app selects the ChatGPT Plan scope, the developer must not enable or encourage creating multiple accounts, splitting usage, rotating accounts, bypassing limits, or “Pooling, transferring, reselling, gifting, or sharing ChatGPT plan usage or Authentication Tokens,” or “Using one user’s subscription to fulfill another user’s requests.” Requests must come from the user’s local runtime or a remote runtime only that user controls. “Use the user’s plan only for the application they connected. Do not provide general-purpose API access for other tools or unrelated requests.” Plan limits still apply, and this sign-in “does not grant extra usage.” OpenAI “may suspend or disable your application’s access” for violations.

OpenAI’s developer docs describe that plan-usage sign-in for “open-source and locally hosted apps,” and tell paid or remotely hosted apps to complete an interest form. The same overview lists “Codex app-server” as a way to use an OAuth access token authorized for ChatGPT plan usage. Source: Sign in with ChatGPT for open source.

The help center, updated two days before the OAuth note was checked, says participating commercial apps can use a Go, Plus, Pro, or Business plan, and “Supported open-source tools remain available to all ChatGPT users.” Eligible requests “use the included ChatGPT Work and Codex allowance or applicable credits.” On a personal plan, apps can use credits after plan limits only if the person has explicitly allowed that, and that switch is off by default. If automatic credit purchases are on, “continued usage can result in automatic charges,” and the person may not get a separate notice when included usage runs out. Source: Using your ChatGPT plan in other apps and sites. These notes did not re-quote Go, Plus, Pro, or Business dollar prices, and the OpenClaw note marks subscription quota numbers as unverified.

OpenClaw’s FAQ says OpenAI “explicitly allows subscription OAuth usage in external tools/workflows like OpenClaw.” That is OpenClaw’s claim. The OpenAI pages above allow the published Sign in with ChatGPT flow, including a local open-source client and a Codex app-server token the user authorized. They forbid the general-API and shared-account shape.

OpenClaw’s Copilot page documents a device login. The person visits a URL and enters a one-time code. OpenClaw stores the GitHub token in a local secret store and then calls that account’s Copilot API. It says the model catalog comes from the same models endpoint VS Code Copilot uses. Model availability “depends on your GitHub plan and organization policy.” The page does not name a minimum paid plan. This chapter does not repeat the login command.

Whose terms apply depends on how Copilot was bought. GitHub’s Additional Products terms, effective 27 August 2026, say Copilot Business and Copilot Enterprise bought directly from GitHub use the GitHub Generative AI Services Terms; purchases through Microsoft use Microsoft’s Product Terms; everyone else uses Section J of the GitHub Terms of Service. Source: GitHub Terms for Additional Products and Features. The Generative AI Services Terms say they apply to volume-licensing customers, and that personal customers who do not buy under a volume license are on the GitHub Terms of Service, including the Additional Products terms.

GitHub’s individual billing page lists Copilot Free, Copilot Pro, Copilot Pro+, and Copilot Max. Published paid allowances: Pro is $10 a month and 1,500 AI credits (1,000 base plus 500 flex); Pro+ is $39 and 7,000 credits; Max is $100 and 20,000 credits. One AI credit equals $0.01. Copilot Free and Copilot Student have an allowance and “access to models through auto model selection only.” Copilot Free includes 2,000 code completions a month. Included credits reset at 00:00:00 UTC on the first day of each calendar month and do not carry over. The same page lists “Third-party coding agents” among features billed in AI credits. GitHub’s separate page of that name is a public preview of Claude and Codex agents inside GitHub, powered by an existing Copilot subscription for Pro, Pro+, and Max. That page does not describe OpenClaw. Sources: Usage-based billing for individuals, Third-party coding agents.

When the included credits run out, GitHub’s individual page gives three choices: upgrade (the charge is the price difference, and earlier usage counts toward the new allowance), stay and set a budget for additional usage, or wait for the next monthly reset. Additional usage “is billed after it occurs.” A person must pay usage incurred before they cancel, drop to Copilot Free, or cut the additional-usage budget to $0. OpenClaw’s Copilot page does not document a separate fallback when that allowance is gone. Business and Enterprise users who exhaust credits see a banner to request more budget from an administrator. No OpenClaw-specific Copilot ban showed up in the pages checked. Organization policy can hide models.

CLIProxyAPI

The project README says CLIProxyAPI “is a proxy server that provides OpenAI/Gemini/Claude/Codex/Grok compatible API interfaces for CLI.” It says a person can reach those providers “locally and with multiple CLI accounts through any OpenAI (including Responses), Gemini (including Interactions), or Claude-compatible client or SDK.” The overview advertises OpenAI Codex via OAuth login, Claude Code via OAuth login, and Grok Build via OAuth login, each including multi-account load balancing where the README says so; “simple CLI authentication flows” for Gemini, OpenAI, Claude, and Grok; and “multiple accounts with round-robin load balancing” for those four. The provider blurb adds that it “supports Gemini API, AI Studio, Vertex AI, Gemini CLI, and Antigravity accounts,” with model availability varying by channel, and that it also accepts ordinary API keys for the Generative Language API. Source: CLIProxyAPI README. Those are the README’s claims. This chapter does not describe how to connect those logins or how to spread traffic across accounts.

If those Gemini or Claude logins are consumer subscriptions, the account risk is the risk above, and the README does not grant a separate permission:

Gemini CLI OAuth or Antigravity OAuth used from this proxy is the third-party access Google’s terms call a breach, with suspension or termination named as the consequence. The February 2026 ban note specifically included proxies, and a second violation is described as permanent. Consumer Gemini CLI Login with Google was already withdrawn on 18 June 2026.
A Claude subscription used so that other clients can call Claude is the pattern Anthropic’s Usage Policy describes as proxying access through a consumer subscription. The login help page says the account holder pays usage-credit charges, and that misrepresenting the client or forcing the traffic onto subscription limits may be enforced. The 15 June 2026 pause, quoted above, still says third-party app usage draws from subscription limits.
Round-robin across multiple accounts is what the README advertises. Google’s reinstatement note treats circumventing usage limits as prohibited. OpenAI’s Sign in with ChatGPT terms forbid rotating accounts, pooling plan usage, and general-purpose API access for other tools. This chapter has no separate count of CLIProxyAPI-specific bans, and no separate Grok terms page.

gpt4free

xtekky/gpt4free is a community project, created by @xtekky, maintained by @hlohaus, and licensed under the GNU General Public License v3.0. The README says the program is provided without any warranty. It describes a local interface, a Python client, a browser client, and a server that speaks an OpenAI-compatible inference API. The local example in the README is http://localhost:1337/v1. Model availability, the README says, depends on what each provider can do that day.

The docs page g4f.dev/docs/ready_to_use.html, as fetched on 14 November 2026, marks these base URLs as no API key required: a local address, and g4f.space paths labeled Groq, Ollama, Pollinations, Nvidia, and Gemini. It marks https://g4f.space/v1 as key required, “many models,” with a key from g4f.dev. A later note on the same page says the hosted URL moved from host.g4f.dev/v1 to g4f.dev/v1. Those two sentences do not name the same host. The same page also lists ordinary vendor APIs beside 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.”

The hosted 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, and 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. If the user picks Auto, “the system picks a working provider automatically.” The operator says they have 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 operator may keep usage logs and excerpts of prompts and responses for up to 14 days. 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. The prompt still leaves the device for whichever provider was chosen.

The README says some adapters need API keys, and some need browser cookies or HAR files because those providers are reached through browser automation. A cookie or HAR file is a logged-in session. This chapter does not describe how to collect one. Pointing personal prompts at a public g4f.space URL means both the proxy operator and an unnamed upstream can read them.

The hosted terms (also version 1.2, 22 October 2026) call the service experimental and “as available.” They say there is no entitlement to a specific model, provider, speed, or quota, and that the operator may discontinue the service at any time without notice. They also forbid circumventing rate limits, bulk automated requests, and reselling access. The relays note says the January 2026 commit history on the repository is full of providers being removed.

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.

The pricing page lists pay as you go at $0 per month, with a balance you top up. It says there is one free model, available on the web only, so that free model is not described as a free API tier. It also lists $12 per month for included usage on open-source models, on the web and through the API. The published allowance is 60 million input tokens per week (cached input counts in full), free output on the included models, 100 images per day on 4 image models, and a reset every Monday at 00:00 UTC. Named families include GLM, DeepSeek, Kimi, MiniMax, MiMo, and Qwen. 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, or choosing a specific provider. Paid models still bill the balance. A 5% discount on eligible paid text models is the page’s claim.

The page says text usage 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.” That no-markup sentence is NanoGPT’s claim. Two opt-in exceptions are printed next to it: manually pinning a specific provider adds 5%, and bring-your-own-key requests are billed 5% of the normal model cost. A “most used models” table that day showed per-million rates such as Gemini 4 Argon at $2 input / $10 output and Claude Opus 5.5 at $4 / $20. Those rows are the page’s table on that day. The page says the full list loads with account pricing.

The pricing banner says “start from $0.10 crypto or $1 card.” The payment FAQ repeats those minimums, says NanoGPT accepts 29 crypto options including Monero, Bitcoin, Nano, USDC, and USDT, plus card through Stripe, and says every crypto deposit gets a 3% bonus. It says crypto “needs no ID or account,” and that card payments are less private because Stripe processes the card. That no-ID sentence is NanoGPT’s claim about its own checkout. Balance is held in US dollars. A subscription is optional, on their page.

nano-gpt.com/provider says NanoGPT auto-routes every request to a provider by default, and that a caller can pin a provider. Pinning is the case the pricing page says adds 5% and excludes from the $12 allowance. The provider-selection docs say the default price is what you pay when you do not select a provider, and that provider prices and availability can change between lookup and the request. Each provider row can include a weight-precision field, such as fp8 or bf16. A floor they describe as fp8 still allows int8, fp8, fp16, bf16, and fp32. The default route can be a more compressed copy than another provider selling the same model name. Their own check, for the most-used models that can be pinned to at least three providers, compares 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. They say it is omitted for models they do not check.

RelaySwitch

relayswitch.io/pricing says every model is priced in USD per million tokens, and that you prepay $10, $25, $50, $100, $250, or $500 by card or by crypto (USDT, USDC, BTC, ETH, and SOL). It says there is no subscription, credit does not expire, and there is no KYC. The no-KYC page says an email address opens the account, that 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 machine that runs a given model still receives the prompt.

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 a base-URL line says 313 models. The proof page header, generated 2026-11-14T13:26:44.569Z, says 358 models checked. The no-KYC page prints a “flagship” table 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. The under-list figures are the site’s comparisons. The maker pages checked the same day list Claude Fable 5.2 at $8 / $40 and Claude Opus 5.5, API id claude-opus-5-5, at $4 / $20, which matches the list column RelaySwitch printed.

The proof page 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 means the model answered the canary the way it did on the earlier runs; Changed means “what serves the id is not what served it before”; Not reached means no route answered the latest run. The snapshot that day: 358 checked, 325 stable, 20 changed, 13 not reached. A stable row means the canary answer matched earlier runs of RelaySwitch’s own prompt. It is not an independent check of the model file. Twenty models were already marked changed on the day this page was opened.

Chutes

chutes.ai/pricing says “Pay per token. No subscription, no minimum, no markup.” The FAQ says there is no free tier, that monthly plans are no longer sold, and that an existing plan runs to the end of the current period and does not renew. Payment is “25+ payment methods via Stripe including crypto, cards, and bank transfers,” plus TAO and Bittensor alpha tokens. The “no markup” sentence is Chutes’ description of its own rates. The page still sets a price per model. It does not claim those rates equal each model maker’s list.

The table is live. One row it showed that day: google/gemma-4-31B-turbo-TEE at $0.12 input and $0.37 output per million tokens. 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.

The privacy page says Chutes does not log, store, or persist API request or response content on Chutes systems, subject to exceptions it lists, and that 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. Confidential-compute models, flagged in the catalog, are described as running inside hardware enclaves on independent operators, with the design meant to keep the machine’s operator from reading the content. Standard models are not in that enclave. The page says many of these are served by third-party inference platforms, and “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.” The page says those are contractual commitments Chutes cannot guarantee. An automatic or fallback route may select a standard model. A confidential model has to be requested by its identifier if every request needs the enclave claim.

Discount relays, with OhMyGPT as an example

Shops in China and elsewhere sell a prepaid key and tell the buyer to point a program at their server, often at a price below the model maker’s list. This is a category, not a recommendation, and this chapter does not describe how to point a program at one.

The usual shape is a prepaid balance. The unit is often a “dollar of credit,” which may not be a dollar of cash. The shop publishes a multiplier or a per-model rate under the vendor’s posted price. The prompt, the files, and the key pass through the shop’s server before they reach whoever actually runs the model. Channels are added and pulled as upstream accounts get rate-limited, shut, or repriced.

apiprice maintains a public directory. 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 that the sticker does not show whether the model received is the model 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: 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.

OhMyGPT (ohmygpt.com, operator named on its own terms as DogeNet LLC) is one long-running shop with a public price page. Listing it 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. The homepage section on being cheaper than the official price says 96 models are below that 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. It is their claim, and it is 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 listed price of $1.50 and $7.50, labeled −10%.

The cost docs say 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.

The privacy policy (last updated 11 September 2025) says OhMyGPT reserves the right to store, review, and investigate communication data, that an upstream such as OpenAI may also record and review it, and that OhMyGPT will not sell interaction inputs and outputs. The terms (last updated 22 October 2023) say the models are supplied by third parties, 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.

Risk table

“Apparent saving” is the gap a flat plan or a discount sticker looks like it offers. Ban risk, surprise-bill risk, and silent-cheaper-model risk are readings of the pages above. “Who sees the prompt” includes the program on the person’s own computer and then every server after it.

Route	Apparent saving	Ban risk	Surprise-bill risk	Who sees the prompt	Silent-cheaper-model risk
Gemini CLI or Antigravity login inside OpenClaw, CLIProxyAPI, or a similar tool	A consumer plan allowance instead of API token charges. Google stopped serving consumer Gemini CLI Login with Google on 18 June 2026.	High. Both terms pages name suspension or termination. February 2026 bans happened, including for proxies, and a second violation is described as permanent.	The published penalty is account action. These pages do not describe an overage bill for this route.	The program on the computer, then Google. Google’s terms call this third-party access.	Low on model identity. The published risk is the account.
Google AI Studio or Vertex API key	No plan arbitrage. The key’s own price applies. This check did not re-quote unpaid quota.	Low under the OAuth rule. Google names this as the supported third-party method.	A paid key bills its Google Cloud project. That bill is the API bill.	The program on the computer, then Google, on the API path Google documents.	Low. The request uses Google’s own API.
Claude subscription login in a third-party app	The monthly plan, if Anthropic still counts the traffic there. The Register used one reporter’s March as an illustration: a $20 plan lined up with about $236 of token usage. That figure is from the 6 April 2026 article.	The Usage Policy forbids proxying through a consumer subscription or misrepresenting the client, and says such use may be enforced. The April 2026 action was a billing change. OpenClaw reports a later staff assurance; that assurance was not on the Anthropic pages in the OAuth note.	Both facts stand. On 15 June 2026 Anthropic paused the announced change and said third-party app usage still draws from subscription limits. The login help page still says Anthropic may draw third-party tool use from usage credits, and the user pays those charges. Bundles bill after plan limits when usage credits are on.	The program on the computer, then Anthropic, from a client other than Claude’s own apps.	Low on model identity when Anthropic is the one answering. A long-context request can require extra usage when the credential is not eligible.
Claude, OpenAI, or Google API key from that company’s own console	None against API rates. This is the price.	Ordinary API enforcement. Anthropic, Google, and OpenAI name an API key as the path for outside programs.	The API bill can be large. It is the bill for the key.	The program on the computer, then the company that issued the key.	Low. The request uses that company’s API.
ChatGPT plan through Sign in with ChatGPT, inside the published rules	Included ChatGPT Work and Codex allowance until the plan limit. Open-source tools are described as available to all ChatGPT users. Commercial plan-usage requires Go, Plus, Pro, or Business.	Low while the app stays inside the published rules. OpenAI can suspend that app’s sign-in access for a violation.	Medium, and off until the person opts in. Credits after the plan limit require an explicit allow. Automatic credit purchases can charge without a separate notice at the limit.	The program on the computer, then OpenAI, on the sign-in path OpenAI documents.	Low. The allowance is OpenAI’s own plan usage. These notes did not re-quote the quota.
Codex or ChatGPT login shared, rotated, or exposed as a general API, the shape the CLIProxyAPI README describes	Several accounts’ allowances, via the README’s round-robin claim.	The Sign in with ChatGPT terms forbid pooling, one person’s plan serving someone else’s requests, and a general-purpose API for other tools. OpenAI may suspend the app’s access. Google and Anthropic rules above apply when the same proxy presents those logins. This check found no separate tally of bans for this project.	Depends on credit and auto-purchase settings on those accounts. The ChatGPT terms do not describe this shape as included usage. Claude usage-credit charges remain the login page’s rule.	Other programs call the proxy, and the proxy presents CLI-account logins. The README says the access is local. The model company still receives the prompt.	The README does not describe swapping the model. The account risk is the one in the Gemini, Claude, and ChatGPT rows.
GitHub Copilot device login in OpenClaw	That account’s included Copilot credits instead of a separate model API bill. Pro is $10 and 1,500 credits, Pro+ is $39 and 7,000, Max is $100 and 20,000. One credit is $0.01. Free is 2,000 completions and auto model selection only.	No OpenClaw-specific Copilot ban showed up in the pages checked. Organization policy can hide models.	Medium once extra usage is allowed. GitHub bills additional credits after the fact. With no budget, the published choices are upgrade, set a budget, or wait until 00:00:00 UTC on the first of the month.	The program on the computer, then GitHub’s Copilot API. OpenClaw says it uses the same models endpoint as VS Code.	Copilot Free publishes auto model selection only, so the model is chosen for you. That is the product, not a hidden swap. Paid plans bill a named credit budget.
gpt4free public endpoints (`g4f.space` and the g4f.dev chat)	$0 on the bases the docs mark as no key. Hosted `/v1` asks for a key.	Upstream sites can forbid this access. g4f’s own terms forbid reselling the service and bulk automated use. Adapters that use a browser cookie put a logged-in session on that path.	The public no-key bases do not show a meter. There is also no promised quota, and the operator may drop the service without notice.	The proxy operator (policy: excerpts up to 14 days) and whichever of 100+ providers received it. Auto picks the provider. That provider may keep the prompt and may train on it.	High. Auto is “a working provider.” The model string sent is not a promise of one company’s weights.
NanoGPT	Pay as you go at their claimed list price, or $12 a month for a capped open-source allowance. Pinning a provider adds 5%. Crypto minimum $0.10, card $1, plus a claimed 3% crypto bonus.	You are the upstream’s customer only through NanoGPT. Closed models sit outside the $12 bundle and keep billing the balance.	The $12 cap is the published open-source allowance. Closed models, which the page names as Claude, GPT, Gemini, and Grok, keep debiting the balance. Prices can change between lookup and the request.	NanoGPT, then the upstream they routed to. Default routing is their choice.	Their docs say weight precision differs by provider, so the default route can be a more compressed copy of the same model name. Their token-count check treats a repeated mismatch as usually a different model, and they do not run it on every model.
RelaySwitch	Prepaid from $10, per token, card or the coins they name. They advertise rates under vendor list. No monthly fee, on their pricing page.	“No KYC” is their signup claim. Resale can still break the model maker’s rules and take the route down.	The charge is the prepaid balance. Under-list rates are their claims and were not checked against the model makers. Catalog counts on their own pages disagree (349, 322, 313, and 358).	Their claim: bodies are not stored, and the account is an email. The machine that runs the model still receives the prompt.	Their daily canary is a self-check. That day, 20 models were marked changed and 13 were not reached, out of 358. “Stable” means the answer matched their earlier canary.
Chutes	No free tier. Top up, then per token. They say no subscription, no minimum, and no markup.	The pages opened do not describe a Chutes-specific account ban. An automatic route can leave the confidential-compute flag.	Pay as you go on a catalog they say is live. The “no markup” line is their claim. A private GPU on that page was $2.25 an hour plus a $6.75 deployment fee, that day.	Their policy: they claim not to persist API content, with listed exceptions. Standard models are sent to third-party platforms whose staff can read the content. The enclave route is a separate, narrower claim.	The bill names a model and a rate. The privacy page says a non-enclave model may be a third-party platform. The “no markup” sentence does not rule out compression or substitution.
Discount relay, category. OhMyGPT is one open example, not an endorsement.	Prepaid credit, often a fraction of list, after a currency rate and a bonus. OhMyGPT’s “below official” badge uses a 7:2:1 blend, and it bills 250,000 credits per dollar with per-line rounding.	Vendor terms often forbid resale and unsupported regions. A ban upstream strands the prepaid balance. The buyer has no contract with the model maker.	OhMyGPT’s terms generally refuse refunds. Their own tiny-request example cost 3 credits against a list of 4. The homepage blend can differ from the share of list on a real call.	The shop, then an upstream you may not see. OhMyGPT’s policy reserves the right to store and review communications, and says the upstream may do the same. Their terms also take a broad license over prompts and generated assets.	apiprice says the sticker does not show that you received that model. OhMyGPT picks the supplier when you do not. Some of its cheap rows put the discount in the model name.

Cost rows

Scores in “How smart it feels” are judgments for this chapter. They are not benchmarks and not vendor scores. A 5 means the route feels like a current flagship from the company that made it, when that company is the one answering. Catches use plain badges: ToS / ban risk, Privacy / logging risk, Surprise bill.

Option	Real monthly cost	Credit card or hardware	How smart it feels (1-5)	Catch
API key from Google AI Studio or Vertex, Claude Console, or OpenAI	Pay as you go. These notes do not measure a light month or a heavy month. The bill is whatever the key uses.	Credit card. No extra hardware.	5, a judgment, when the key calls a current flagship from that company. A smaller model on the same key feels lower.	Surprise bill. The bill is the API bill, and a busy agent can spend more than a chat subscription feels like it should.
Gemini CLI or Antigravity login inside a third-party tool	No new consumer login to price. Google stopped serving consumer Gemini CLI Login with Google on 18 June 2026. These notes do not re-quote Google AI Pro or Ultra prices.	Credit card, on the Google plan, if a login still exists. OpenClaw will not create one.	4, a judgment: the named Google models, on a login that still answers.	ToS / ban risk. Google’s terms name suspension or termination. February 2026 bans included proxies, and a second violation is described as permanent.
Claude subscription login inside a third-party app	The plan price, for as long as the 15 June 2026 pause holds, plus usage-credit charges if Anthropic draws the traffic that way. Bundle examples checked the same day: $45 for $50 of credit, $200 for $250, $700 for $1000, then standard rates past a monthly cap. The Register’s $20-versus-about-$236 line is one reporter’s March, in the 6 April 2026 article, not a current rate card.	Credit card.	5, a judgment, when Anthropic is the one answering.	Surprise bill. On 15 June 2026 Anthropic said third-party app usage still draws from subscription limits. The login page still says Anthropic may draw that use from usage credits and that you pay. ToS / ban risk. The Usage Policy forbids proxying a consumer subscription, and says such use may be enforced.
Sign in with ChatGPT, inside OpenAI’s published rules	The ChatGPT plan. These notes did not re-quote Go, Plus, Pro, or Business prices. Included allowance until the limit. Credits after that only if you allow them.	Credit card. Storage has to stay on a machine you control.	4, a judgment: ChatGPT plan models inside the published allowance. Quota numbers were not on the pages opened.	Surprise bill, if automatic credit purchases are on. OpenAI says that charge can arrive with no separate notice when included usage runs out.
Several Claude, Gemini, or ChatGPT logins behind a local proxy, the shape the CLIProxyAPI README describes	The subscriptions on those accounts, until an account is closed. The README’s round-robin line is the README’s claim. No separate proxy fee was on that page.	Credit card on each account. The README says the proxy is local.	4, a judgment: as smart as the account it presents, for as long as that account still works.	ToS / ban risk. Google names suspension for this kind of access. Anthropic’s login page can bill usage credits and may enforce a misrepresented client. OpenAI may suspend an app that pools plans or offers a general API.
GitHub Copilot Free, device login	$0 for Copilot Free, with 2,000 code completions a month and auto model selection only. Completions do not carry over.	No card for Copilot Free. No extra hardware.	2, a judgment: auto model selection only, and a small monthly completion allowance.	The pages checked show no OpenClaw-specific ban. The published free path chooses the model for you.
GitHub Copilot Pro, Pro+, or Max, device login	Pro $10 and 1,500 credits, Pro+ $39 and 7,000, Max $100 and 20,000. One credit is $0.01. Credits reset at 00:00:00 UTC on the first of the month and do not carry over. Additional usage bills after it occurs if a budget is set.	Credit card.	4, a judgment: paid Copilot models, until the credit budget runs out. The pages checked do not list every model.	Surprise bill, once extra usage is allowed. GitHub bills that usage after the fact. With the budget at $0, the published choices are upgrade or wait for the monthly reset.
gpt4free public addresses	$0 on the bases the docs mark as no key. No quota is promised. The hosted path that asks for a key has no price in these notes.	No card for the no-key public addresses. No extra hardware.	2, a judgment: whoever answered that request. Auto means a working provider, not one company’s flagship.	Privacy / logging risk. The operator may keep excerpts for 14 days, and the upstream may store the prompt and may train on it. ToS / ban risk if an adapter is using a logged-in browser session.
NanoGPT	$12 a month for the published open-source allowance (60 million input tokens a week, output free on those models, 100 images a day, reset Monday 00:00 UTC), or pay as you go. Closed models bill the balance. They claim list price with no markup. Pinning a provider adds 5%. Card minimum $1, crypto minimum $0.10, plus a claimed 3% crypto bonus.	Card or crypto. No extra hardware.	3, a judgment, on the $12 open-source shelf. A named closed model would feel higher only if that provider is actually serving it.	Privacy / logging risk. NanoGPT and the upstream it picks can read the prompt. Their docs also say a provider can serve a more compressed copy of the same model name.
RelaySwitch	No monthly fee on their pricing page. Prepaid $10, $25, $50, $100, $250, or $500, then per token. Under-list rates are their claims.	Card or crypto (USDT, USDC, BTC, ETH, SOL). They claim no identity check. No extra hardware.	3, a judgment. The advertised names look like flagships. Their own check the same day marked 20 models changed and 13 not reached.	Privacy / logging risk. They claim not to store bodies. The machine that runs the model still receives the prompt. Their canary is their own test.
Chutes	No free tier and no new monthly plan. Top up, then per token, or per second on their GPU. One GPU row that day was $2.25 an hour plus a $6.75 deployment fee. They claim no markup.	Card, crypto, or bank transfer through Stripe, or TAO. The optional GPU is theirs, billed by the second.	3, a judgment. The catalog rows opened that day were `google/gemma-4-31B-turbo-TEE` and `Qwen/Qwen3.8-27B-TEE`, not a Claude or GPT subscription.	Privacy / logging risk on standard models. Their page says a third-party platform’s staff can read that content. The enclave route is a narrower claim, and an automatic route can leave it.
Discount relay. OhMyGPT is an example, not an endorsement.	Prepaid credits, not a flat fee in the pages opened. OhMyGPT bills 250,000 credits per dollar and rounds each line up. Their “below official” badge is a 7:2:1 blend, which is their claim. A tiny request in their own example cost 3 credits against a list of 4.	Prepaid balance. These notes do not re-state OhMyGPT’s card-versus-crypto line. No extra hardware.	3, a judgment. The shop picks the supplier, and some cheap rows put the discount in the model name.	ToS / ban risk. Resale can close the upstream account and strand the balance, and their terms generally refuse refunds. Privacy / logging risk. They reserve the right to store and review prompts, and their terms take a broad license over prompts and generated assets.

What the layman guide should say

Use an API key from the company that made the model, and pay that company. A Gemini CLI or Antigravity login inside a third-party tool can close the Google account: Google’s terms name suspension or termination, Google suspended accounts for third-party tools and proxies in February 2026, consumer Login with Google for Gemini CLI stopped on 18 June 2026, and OpenClaw no longer creates that login. On 15 June 2026 Anthropic paused an announced change and said third-party app usage still draws from subscription limits. The Claude login page still says Anthropic may draw third-party tool use from usage credits and that you pay those charges. Sign in with ChatGPT can spend the included allowance inside OpenAI’s published rules, and automatic credit purchases can add a charge with no separate notice. Pooling logins, or a proxy that round-robins several Claude, Gemini, or ChatGPT accounts the way CLIProxyAPI’s readme describes, can get the account or the app cut off. A Copilot login spends that GitHub account’s credits and can bill more after the fact once extra usage is allowed. Public aggregators and discount relays, including gpt4free’s public addresses, NanoGPT, RelaySwitch, Chutes, and shops such as OhMyGPT, can read the prompt and can answer from a different or more compressed model. This chapter does not say how to turn any of those routes on.

Sources

Checked 14 November 2026 unless a page date is given above.