Published Updated
In one or two sentences, before anything else.
ZenMux is an LLM gateway, operated under its terms by NexaMind Singapore Pte. Ltd., that sells Claude, GPT, Gemini and many Chinese models through OpenAI, Anthropic and Vertex-style endpoints at the makers' list token prices. It adds automatic credits for bad or slow answers, which it calls insurance, and 20 to 200 USD monthly coding plans.
What it can do, one sentence each, and whether we actually tested it.
Pay-as-you-go token prices equal the makers' list prices, such as Claude Opus 5.5 at 4 / 20 USD and GPT-6 Sol at 2 / 10 USD per million tokens.
A 5% top-up service fee is waived under a promotion that may end without notice.
OpenAI Chat Completions and Responses, Anthropic Messages and Vertex AI / Gemini endpoints, with any model callable through any of them.
Per ZenMux's quickstart and API overview.
Routing prefers the model's own developer and fails over (for Claude: Anthropic, Vertex AI, Amazon Bedrock); callers can pin providers, set model fallbacks and per-key RPM, TPM and spend limits.
Per ZenMux's provider-routing docs; pinned routes can still fail, as a September 2026 Bedrock issue shows.
ZenMux says it credits requests it judges to have poor output, hallucination or high latency the next day, with no claim to file.
No thresholds are published; its counter showed 10,649 USD paid in total on October 21, 2026, and it requires request data to be stored.
The ZenMux/skills repository adds skills for docs questions, tool setup, usage and balance, feedback, a Claude Code status line and image generation, installed with npx skills add ZenMux/skills.
Per ZenMux's docs and GitHub.
20, 100 or 200 USD a month for personal, non-production use, with rolling 5-hour and 7-day quotas and 10 to 15 requests a minute.
Ultra was worth at most 480.40 USD of usage a month on October 21, 2026; the Flow value floats.
Turning off Data Services stops ZenMux storing request content.
Insurance claims stop working in that mode.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
export ANTHROPIC_BASE_URL="https://zenmux.ai/api/anthropic", ANTHROPIC_AUTH_TOKEN to a ZenMux key (subscription sk-ss-v1 or pay-as-you-go sk-ai-v1) and ANTHROPIC_API_KEY="", then run claude.
What happened
ZenMux accepts Anthropic Messages, so Claude Code needs no other change; pay-as-you-go keys bill at Anthropic's list prices. The guide says subscription keys must not be used in production, and Claude routes failed for users on July 27, 2026.
Case 2
Input
Repeated calls to Claude Opus 4.6, then Claude Haiku 4.5, pinned to the amazon-bedrock provider.
What happened
The route kept failing from September 18 to at least September 25, 2026, the model page showed no cache, latency or throughput data, and ZenMux had not replied that week. One user's repeated test of one route.
Case 3
Input
Two identical requests with cache_control to claude-opus-4.6 and to claude-sonnet-4.5, on both endpoints.
What happened
Sonnet 4.5 cached 1,816 tokens on the second call; Opus 4.6 showed zero cache writes and zero cached tokens on both calls. The issue was closed four days later; one model at one time.
Case 4
Input
Coding in OpenCode on the 20 USD Builder Plan with Claude Opus.
What happened
The poster says one Opus request costing about 0.80 USD used 10% of the weekly quota, and a refund after usage left almost nothing; a replier said one task used half of the 200 USD plan's 5-hour window. Quotas have changed since.
What it does badly, with the evidence, and what we could not settle.
Weak
Claude access has broken at the upstream.
Evidence: On July 27, 2026 all Claude routes returned an error saying the organization had been disabled while GPT and GLM worked (GitHub issue #267); the Bedrock route for Claude failed for at least a week from September 18, 2026 (issue #297).
Insurance terms are vague.
Evidence: ZenMux publishes no thresholds or payout formula, and its makers told Product Hunt commenters the thresholds are still being tuned. Its counter showed 10,649 USD paid in total against about 11.5 trillion tokens.
Refunds only within 24 hours.
Evidence: Its refund guide allows requests within 24 hours of a top-up or subscription, deducts usage and a 5% fee, and forfeits bonus credits.
The privacy policy allows training on anonymized content.
Evidence: It says ZenMux "may store, transfer or train on your anonymized User Content" and may share data with affiliates; SOC 2, ISO 27001 and GDPR are all "In progress" while footers show the badges.
Billing errors.
Evidence: ZenMux admitted double-billing cached tokens on seven Gemini models and credited users (March 10, 2026); an issue that cancelled streaming requests are billed in full has been open since April 14, 2026.
No status page.
Evidence: status.zenmux.ai did not resolve on October 21, 2026; model pages show per-provider latency and uptime only.
Not for production systems that depend on Claude, or buyers who need disclosed ownership.
Evidence: Claude routes failed twice in 2026, the subscription is for personal non-production use only, and no page states who owns the operator.
Uncertain
Who owns ZenMux.
Why unsure: The terms name NexaMind Singapore Pte. Ltd.; a March 2026 notice and the home-page footer name AI Force Singapore Pte. Ltd.; the domain's registrant is Advanced New Technologies Co., Ltd., an Ant Group IP company. No page states who owns NexaMind.
Whether ZenMux still serves Claude directly from Anthropic.
Why unsure: After the July 2026 organization-disabled errors, ZenMux has not said whether Claude runs through Anthropic or only through Bedrock and Vertex AI.
Whether bring-your-own-key is live.
Why unsure: The terms describe a BYOK option, but a May 2026 GitHub request asked for it.
Which accounts can use remote MCP tools.
Why unsure: The Responses reference lists the mcp tool type "if enabled" without saying for which models or accounts.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | NexaMind Singapore Pte. Ltd., 11 Collyer Quay, Singapore (terms effective September 9, 2026) | The home-page footer and a March 2026 notice name AI Force Singapore Pte. Ltd. Source: zenmux.aichecked |
| Domain registrant | Advanced New Technologies Co., Ltd., Cayman Islands; zenmux.ai created July 31, 2025 | An Ant Group IP company; contact address on antfin.com. Source: rdap.identitydigital.serviceschecked |
| Price vs makers' list prices | Same as list (Claude Opus 5.5 4 / 20 USD, GPT-6 Sol 2 / 10 USD per 1M tokens) | Source: zenmux.aichecked |
| Platform fee | 5% top-up service fee, waived under a limited-time promotion | Source: zenmux.aichecked |
| Minimum top up and payment | 5 USD (up to 25,000 USD); Stripe cards, Alipay, gift cards | Source: zenmux.aichecked |
| Plans | Free (Studio Chat on the web, no API); Builder Plan 20 / 100 / 200 USD a month for personal, non-production use | Source: zenmux.aichecked |
| Models | 206 in the public model list | Source: zenmux.aichecked |
| API formats | OpenAI Chat Completions and Responses, Anthropic Messages, Vertex AI / Gemini; tool calls, structured output, prompt cache, web search, reasoning | Source: zenmux.aichecked |
| MCP | No official MCP server found in its docs; Responses reference lists remote MCP tools "if enabled" | Source: zenmux.aichecked |
| Agent Skills | Official ZenMux/skills repository, installed with npx skills add ZenMux/skills | Source: zenmux.aichecked |
| Coding agents | Guides for Claude Code, Codex, Cursor, Cline, Gemini CLI, GitHub Copilot, OpenCode, Pi and others | Source: zenmux.aichecked |
| Privacy and data retention | Request content stored by default; optional zero-retention mode; usage records 180 days; may train on anonymized content | Source: zenmux.aichecked |
| Status page | None found; status.zenmux.ai did not resolve | Source: zenmux.aichecked |
| Refunds | Within 24 hours of a top-up or subscription, minus usage and a 5% fee; bonus credits forfeited | Source: zenmux.aichecked |
Answer first, then why, an example, the limits and what to do next.
ZenMux charges the same token prices as the official APIs on pay-as-you-go, not less. Claude Opus 5.5 is 4 / 20 USD per million tokens and GPT-6 Sol 2 / 10 USD, as at Anthropic and OpenAI, and the 5% top-up fee is waived for now.
Limits
The Builder Plan can cost less per token for personal coding, but its quotas are capped and the Flow value floats.
ZenMux is a working gateway with a public issue tracker, but its ownership is not stated and its privacy policy allows training on anonymized content. Its terms name NexaMind Singapore Pte. Ltd., its domain is registered to Advanced New Technologies Co., Ltd., an Ant Group IP company, and its SOC 2, ISO 27001 and GDPR work is marked "In progress".
Limits
Turning off Data Services stops request storage but also ends insurance credits.
ZenMux's free plan covers only Studio Chat on the web, with 5 Flows every 5 hours and no API access. That is roughly five conversations per window by ZenMux's estimate; a V2EX user reported one long Opus chat used the whole window.
Claude Code connects to ZenMux by setting ANTHROPIC_BASE_URL to https://zenmux.ai/api/anthropic and ANTHROPIC_AUTH_TOKEN to a ZenMux key. ZenMux's guide also clears ANTHROPIC_API_KEY and says subscription keys (sk-ss-v1) must not be used in production; pay-as-you-go keys start with sk-ai-v1.
ZenMux publishes official Agent Skills but no official MCP server. Its ZenMux/skills repository installs with npx skills add ZenMux/skills and covers docs questions, tool setup, usage and balance, feedback, a Claude Code status line and image generation; its Responses API reference lists remote MCP tools "if enabled".
ZenMux costs slightly less than OpenRouter for now, because both bill at the makers' list prices and only OpenRouter currently adds a fee. OpenRouter charges 5.5% on credit purchases; ZenMux's 5% top-up fee is waived under a promotion that can end without notice.
Limits
ZenMux had two reported Claude route failures in 2026 and refunds only within 24 hours.
This page is built from ZenMux's own site, docs, terms, model list and GitHub, the domain registry, the official price pages of OpenAI, Anthropic and Google, and user reports on GitHub, V2EX, Hacker News and Product Hunt, all checked on October 21, 2026. ZenMux's claims are labelled as its own.
ZenMux is an LLM gateway with three native request formats: OpenAI (https://zenmux.ai/api/v1, Chat Completions and Responses), Anthropic Messages (https://zenmux.ai/api/anthropic) and Vertex AI / Gemini (https://zenmux.ai/api/vertex-ai). Its quickstart says any model can be called through any of them. Its public model list held 206 models on October 21, 2026, weighted toward Chinese makers such as Qwen and Z.AI, next to Claude, GPT, Gemini, Grok and DeepSeek. Default routing prefers the model's own developer and fails over to other providers; for Claude these are Anthropic, Vertex AI and Amazon Bedrock. Callers can pin providers, set model fallbacks and give each key RPM, TPM and spend limits.
It sells two ways. Pay-as-you-go tops up from 5 USD to 25,000 USD by Stripe card or Alipay. The Builder Plan is a monthly subscription for personal, non-production coding, billed in "Flows" whose dollar value floats:
| Plan | Price a month | Quota | Most usage a month, per ZenMux's pricing page on October 21, 2026 |
|---|---|---|---|
| Free | 0 | 5 Flows per 5 hours, Studio Chat on the web only | 5.44 USD |
| Starter | 20 USD | 50 Flows per 5 hours | 30.01 USD |
| Max | 100 USD | 300 Flows per 5 hours | 180.15 USD |
| Ultra | 200 USD | 800 Flows per 5 hours | 480.40 USD |
Plans allow 10 to 15 requests a minute, and ZenMux may lower the Flow value for accounts it judges abusive. Its headline extra is "AI insurance": ZenMux says it credits requests it judges to have poor output, hallucination or high latency the next day, with no claim to file.
Pay-as-you-go token prices match the makers' list prices for every model checked:
| Model, per million input / output tokens | ZenMux | Maker's list price | Source |
|---|---|---|---|
| Claude Opus 5.5 | 4 / 20 USD | 4 / 20 USD | Anthropic |
| Claude Sonnet 5.5 | 2 / 10 USD | 2 / 10 USD | Anthropic |
| GPT-6 Sol, up to 272K input tokens | 2 / 10 USD | 2 / 10 USD | OpenAI |
| GPT-5.4, up to 272K input tokens | 2.50 / 15 USD | 2.50 / 15 USD | OpenAI |
| Gemini 4 Argon | 2 / 10 USD | 2 / 10 USD until November 30, 2026; 4 / 20 USD from December 1 |
ZenMux prices come from its public model list. The margin is a 5% top-up service fee, which its promotions page says is waived for a limited time and returns when the promotion ends. Some providers of the same model are priced differently, and a few Chinese models have time-of-day pricing.
ZenMux has no public status page (status.zenmux.ai did not resolve), and no third-party latency or throughput benchmark of it was found. What exists:
cache_control, Claude Sonnet 4.5 cached 1,816 tokens on the second call while Claude Opus 4.6 cached none on either call. Closed four days later.On July 27, 2026 all Claude routes returned an error saying the organization had been disabled, while GPT and GLM kept working (GitHub issue #267, closed the same day). In March 2026 ZenMux admitted double-billing cached tokens on seven Gemini models and credited users. An issue that streaming requests cancelled by the client are billed in full has been open since April 14, 2026.
A 43-reply V2EX thread from March 2026 is mostly negative about Builder Plan quotas, refunds after usage and slow email support; quotas have changed since. On Product Hunt (third of the day on February 13, 2026, no written reviews), the makers said insurance thresholds are still being tuned. Trustpilot has one one-star review.
No official MCP server was found in ZenMux's docs. Its Responses API reference lists remote MCP tools (type mcp, with server URL and allowed tools) "if enabled", and its tool-calling guide says Responses requests can carry MCP tools.
ZenMux does publish official Agent Skills: the ZenMux/skills repository, installed with npx skills add ZenMux/skills. Its skills answer questions from ZenMux's docs (zenmux-context), walk through tool setup (zenmux-setup), read quota, balance and per-request cost through the Management API (zenmux-usage), file GitHub issues (zenmux-feedback), add a Claude Code status line (zenmux-statusline) and generate images (zenmux-image-generation), plus two smaller ones. Anthropic's Skills API through its Messages endpoint is not documented.
Its docs have setup guides for Claude Code, Claude Desktop, Codex, Cursor, Cline, Gemini CLI, GitHub Copilot, OpenCode, Pi, Neovate Code, DeepSeek Harness, Hermes Agent and OpenClaw. Its GitHub has OAuth sign-in provider plugins for Codex, opencode, Pi and Hermes Agent. For Claude Code, you set ANTHROPIC_BASE_URL to https://zenmux.ai/api/anthropic and ANTHROPIC_AUTH_TOKEN to a ZenMux key; the guide says subscription keys must not be used in production. Code uses the OpenAI, Anthropic and Google GenAI SDKs as they are, and the docs have an llms.txt index.
OpenRouter also bills at list prices, plus 5.5% on credit purchases, and has more independent coverage. NanoGPT charges list prices with no deposit fee and has a 12 USD plan for open-weight models. CometAPI sells at 0.8 times list but names no company, and AIMLAPI charges about 1.3 times list. ZenMux's extras are the insurance credits, official Agent Skills and the coding subscription.
Anyone whose product depends on steady Claude access, anyone who needs to know who owns their supplier, and anyone who wants refunds after the first day. The Builder Plan is not for production use under ZenMux's own rules.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Requesty is a London-based LLM gateway, run by Requesty Ltd and seed-funded by 20VC in 2025, that sells OpenAI- and Anthropic-compatible access to more than 600 models at list price plus a 5% top-up fee. It stands out for routing policies, spend controls and an EU endpoint in Frankfurt.
Poe API is Quora's developer API for the bots in its Poe chat app, opened on July 31, 2025, with OpenAI-compatible endpoints and an Anthropic-compatible one for official Claude bots. Calls spend the same points as a Poe subscription, so there is no separate API bill; current Claude bots cost about 1% more than Anthropic's list price.
OpenRouter is a New York company's API relay that sells access to more than 680 models from over 100 hosting providers through one key, at each provider's list price plus a 5.5% fee on credit purchases. Stripe agreed to buy it on August 19, 2026; no closing had been announced by October 21, 2026.
OpenCode Zen, called OpenCode Console in the docs since October 2026, is the pay-as-you-go model gateway from Anomaly Innovations, the team behind the OpenCode coding agent. It sells about 80 coding models near the makers' list prices plus a 4.6% + 0.31 USD card fee; a Console breach that month may have exposed some users' saved provider keys.