This page is built from CometAPI's own site, docs, terms, status page, public model catalogue and GitHub organisation, the domain registry, the official price pages of OpenAI, Anthropic and Google, GitHub issues, Product Hunt and Hacker News, all checked on October 21, 2026. CometAPI's claims are labelled as its own.
CometAPI sells one prepaid key for text, image, video and audio models from many makers. Its marketing says "500+" models; its public catalogue returned 322 on October 21, 2026 (116 text, 42 video, 40 image, 20 audio, the rest untyped). They include GPT-6 Sol, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7, DeepSeek V4.1, Kimi K3, Sora 2, Veo 3.1, Kling, Runway, FLUX 2 and 42 Midjourney entries. The catalogue also lists models the makers have not released, such as claude-opus-6, marked "Coming Soon" and priced at a placeholder 75 / 75 USD.
New accounts get 0.1 USD of credit with no card. Top-ups start at 10 USD by card through Stripe, the only documented payment method, and credits do not expire while the account is in good standing. When an upstream channel fails, CometAPI's Help Center says the system switches channels and retries; callers cannot choose which upstream serves a request.
Prices against the official APIs
CometAPI's pricing page states the rule "Customer price = official x 0.8" for official models (Kling endpoints use 0.95). Its catalogue lists each model's official price and a 0.8 ratio. The CometAPI column below applies that rule to the catalogue:
| Model, per million input / output tokens | CometAPI | Maker's list price | Source |
|---|
| Claude Opus 5.5 | 3.20 / 16 USD | 4 / 20 USD | Anthropic |
| Claude Opus 5 | 4 / 20 USD | 5 / 25 USD | Anthropic |
| Claude Sonnet 5.5 | 1.60 / 8 USD | 2 / 10 USD | Anthropic |
| GPT-6 Sol, up to 272K input tokens | 1.60 / 8 USD | 2 / 10 USD | OpenAI |
| Gemini 4 Argon | 1.60 / 8 USD | 2 / 10 USD until November 30, 2026; 4 / 20 USD from December 1 | Google |
CometAPI explains the discount as an "intelligent routing system", buying API credits in bulk from several providers and servers in several regions. Its docs do not agree with themselves: one page says pay-as-you-go is "consistent with official pricing", another gives the 0.8:1 ratio, and the Help Center describes per-model "multipliers" that do not map directly to official prices and change with market conditions for non-GPT models. Midjourney, Kling, Luma and other products without official APIs are priced by CometAPI itself, which says it implements them "through technical means".
What public tests show
No third-party latency, throughput or output-quality test of CometAPI was found in public sources. What exists:
- CometAPI's own model status page (Uptime Kuma probes of 56 models, live on October 21, 2026) showed 24-hour availability of 93.9% for claude-opus-5, 94.5% for claude-opus-4-7, 95.3% for gpt-5.6 and 97.3% for gpt-6-sol. Gemini and DeepSeek probes were at 98% to 99%, the median was 97.9% and the lowest 60.6% (grok-build-0.1). The method is not documented.
- A GitHub user in the Hermes Agent tracker (September 25, 2026) called claude-opus-5-5-thinking through CometAPI and got an HTTP 400 passed through from AWS Bedrock Runtime InvokeModel. The log shows the request was served through AWS Bedrock and that CometAPI did not convert the old thinking parameter to the adaptive form the model requires.
- CometAPI's own latency benchmark page defines time to first token and throughput; its sample figures (640 ms, 42.1 tokens a second) are labelled as an illustrative format, not measured CometAPI results.
- A CometAPI blog post compares CometAPI with OpenRouter and AIML API; it is a competitor comparison written by the vendor and contains no measurements.
What users report
Independent reviews barely exist. Trustpilot has no reviews for cometapi.com (the unclaimed profile); negative reviews that turn up for "Comet" are about cometgift.com, an unrelated retailer. Product Hunt shows 6 points, 12 followers, no reviews and one comment, from the maker. G2 shows about two reviews in search results. CometAPI's pricing page nevertheless carries badges of 5.0 out of 5 on Product Hunt and 4.9 out of 5 on G2.
Most visible activity is promotion. On Hacker News, an account named cometapi posted 61 items and other accounts posted identical marketing text. On GitHub, an account opened issues on more than 20 open-source projects in September and October 2025 (the Vercel AI SDK, Aider, Milvus and others) asking them to add CometAPI as a provider, and an account named CometAPI-Official continues in 2026. A press release on GPT-6 Astra went out through GetNews in September 2026.
MCP, Agent Skills and coding agents
No official MCP server was found in CometAPI's docs or on its GitHub organisation (cometapi-dev). Its Responses API docs cover function calling and the built-in web_search tool and refer to OpenAI's tools guide for other built-in tools; passing a remote MCP server through a request is not documented. Its Messages API docs cover prompt caching and the server tools web_search and web_fetch, but no MCP connector and no Anthropic Skills. No official Agent Skills were found either.
For coding agents, the docs list more than 50 integration guides, including Claude Code (set ANTHROPIC_BASE_URL=https://api.cometapi.com and ANTHROPIC_AUTH_TOKEN), Codex, Gemini CLI, OpenCode, Cline, Continue, avante.nvim, gptme, Hermes Agent, OpenClaw and Pi. Cursor is not among them. A desktop helper, CometAPI Connect, writes CometAPI settings into installed tools; its README said the v0.1.3 release was private while CometAPI completes a publication review. For code, there is a Vercel AI SDK provider (@cometapi/ai-sdk-provider), official Python and Node SDK repositories, guides for LiteLLM, LangChain and LlamaIndex, and an llms.txt index of the docs.
Privacy, terms and who runs it
CometAPI's terms, privacy policy, refund policy, SLA, Trust Center and About page name no legal entity and no address; the footer carries only a CometAPI copyright line. The terms choose Hong Kong law with exclusive jurisdiction of the Hong Kong courts. The domain cometapi.com was registered on July 2, 2024, and its first Hacker News posts date from January 2025. The Help Center uses the vocabulary of the open-source One API and New API relay software: channels with automatic switching, group and completion multipliers, and a Tokens menu for API keys.
- CometAPI says it does not collect or store prompts, inputs or outputs, and records IP addresses, request frequency and quota use. The Trust Center says request data is not used for training and "Formal certifications have not yet been initiated"; the About page calls SOC 2 Type II a goal.
- There is no GDPR section, data region, list of upstream providers or named data controller.
- Credits are generally non-refundable; unused credits may be refunded at CometAPI's discretion, and disputes must be raised within 30 days. Gateway rejections and upstream 5xx errors are not billed.
- The SLA targets 99.9% monthly uptime for the gateway only, excludes upstream disruptions and names no service credit. A daily maintenance window runs from 1 to 5 AM, time zone not stated.
How it compares with other relays
| Service | Price against makers' list prices | Who runs it, per its own terms |
|---|
| CometAPI | 0.8 times | No company named; Hong Kong law |
| OpenRouter | Same, plus 5.5% on credit purchases | OpenRouter, Inc. |
| AIMLAPI | About 1.3 times | Boiler Labs FZ-LLC; AIMLAPI OÜ |
| ZenMux | Same, 5% top-up fee waived for now | NexaMind Singapore Pte. Ltd. |
| Kie.ai | Below list for many models | Two LLCs at a Denver address |
CometAPI is the cheapest of these for Claude, GPT and Gemini on paper. The others either charge list price or more, but say who you are contracting with, and OpenRouter lets callers choose providers.
Who should skip it
Businesses that must know their supplier, need certified data handling or data-residency terms, or need steady Claude and GPT access for production work. For personal projects where a 20% saving matters more than an anonymous operator, the risk is bounded by the credit you top up.