codexhot.com

Codexhot (热核算力, "hot-core computing power") is a paid proxy for OpenAI's GPT API, not a multi-vendor aggregator. Its own homepage meta description states the terms in one line: "付费 GPT API" (paid GPT API) with GO/PLUS/PRO billed at 2%/7%/10% of the site's own model base rate, vouchers starting at RMB 5 (https://www.codexhot.com/, checked 2026-10-09). Its Messages-API guide goes further and explicitly disclaims the obvious misreading: offering an Anthropic-shaped /v1/messages endpoint "does not mean Claude models are available" here (https://www.codexhot.com/guides/messages/). Every model listed across its docs — gpt-6-luna, gpt-6-astra, gpt-6-sol, gpt-5.6-luna/sol/terra — matches OpenAI's own current model catalog (https://developers.openai.com/api/docs/models), so the model names are real; what's being resold is access to one vendor's API, under Codexhot's own pricing, not a connection to "multiple AI model providers."

There is no free tier or trial. To use the API you register, buy a prepaid redemption code (兑换券, starting at RMB 5) from a third-party code shop linked from Codexhot's own guides (https://wzyp.cn/shop/UPUD9X5R), redeem it on Codexhot for account balance, then create an API key bound to one of three groups. GO (0.02x / 2%) currently whitelists 3 models, PLUS (0.07x / 7%) whitelists 8, and PRO (0.1x / 10%) whitelists 1 (https://www.codexhot.com/guides/groups/). Those percentages are explicitly against Codexhot's own "本站模型基础计费" (this site's model base billing) — the groups guide says so directly: "比较对象是本站模型基础计费，不是其他服务商的售价" ("what's being compared is this site's own model base billing, not another provider's price"). A worked example on the same page: gpt-6-luna plain text on GO runs about $0.002 per million input tokens and $0.01 per million output tokens, in the site's own billing units — not a quoted discount off OpenAI's published per-token price, which Codexhot doesn't reference anywhere in its pricing math.

Codexhot's guides are unusually specific for a reseller in this category. They give working curl, Python and Node.js examples for three different wire protocols at once — OpenAI's Responses API, OpenAI's Chat Completions API, and an Anthropic-shaped Messages API — each pointed at the same /v1 base URL (https://www.codexhot.com/guides/gpt-api/, /guides/sdk/, /guides/chat-completions/). The error-handling guide lists concrete error codes and bodies to check (API_KEY_QUOTA_EXHAUSTED, gateway_concurrency_limit, model_not_found) rather than generic troubleshooting advice, and includes a self-check request (GET /v1/models) that doesn't burn inference quota (https://www.codexhot.com/guides/errors/). The groups guide is also candid about its own limits: it says a model is only added to a tier's public whitelist after real request testing, and warns readers not to assume the PRO name implies extra performance ("不要仅凭'PRO'名称推断还有其他模型或更高响应质量").

Codexhot's error guide says a 503 means "no dispatchable account was currently available" to serve a request — evidence the backend pools multiple upstream accounts, but the guide stops there: it offers retry-later advice, not a description of an automatic-failover feature a customer can configure or rely on (https://www.codexhot.com/guides/errors/). Its sitemap lists only the homepage and seven guide pages — no about, terms, or company page (https://www.codexhot.com/sitemap.xml) — and the only support channel named anywhere in its docs is a QQ group ("官网客服 QQ 群：1106148284"); no support email or ticket system is published. There's no changelog, status page, or security/compliance page. OpenAI's own Business Terms prohibit reselling or sharing access to an OpenAI account (https://openai.com/policies/business-terms/); nothing on Codexhot's site states how it sources the GPT access it resells or whether it operates under an authorized arrangement, and a search turned up no independent coverage, reviews, or complaints about Codexhot to corroborate either way.