Tokeness review: one API key for GPT/Claude/Gemini/Grok/DeepSeek/Kimi (with real caveats)

A developer consolidated multi-model API usage onto Tokeness, a third-party API gateway that provides a single key for GPT, Claude, Gemini, Grok, DeepSeek, GLM, and Kimi. The gateway offers transparent per-1M-token pricing, a usable free tier, and business invoices, but has caveats including blocked access from mainland China for GPT/Claude and the inherent trust issues of a relay service.

TL;DR: Moved my multi-model API usage to Tokeness tokeness.io , a third-party API gateway. One key covers GPT, Claude, Gemini, Grok, DeepSeek, GLM and Kimi; prepaid CNY billing 7:1 with business invoices; per-1M-token prices listed openly in the console; genuinely usable free tier. Main caveats: GPT/Claude are blocked for mainland China IPs their own compliance decision, not a bug , it's a relay so your prompts pass through their servers, and prices aren't locked. Full breakdown below. Background I run Claude Code and Codex CLI daily, and was maintaining separate accounts for OpenAI, Anthropic and DeepSeek — three billing systems, three key dashboards, plus payment friction without an international card. A few weeks ago I consolidated onto Tokeness. Not affiliated with them; just sharing what I found, since multi-model gateways get discussed here a lot. What it is A relay/gateway: you point your tools at https://n.tokeness.io/v1 https://n.tokeness.io/v1 , they forward to the official APIs. Both OpenAI-compatible and native Anthropic protocols are supported, so Claude Code, Codex CLI, Cursor, Cline, Dify, n8n etc. are a one-line base url change they have ~20 integration guides at docs.tokeness.io . from openai import OpenAI client = OpenAI api key="YOUR TOKENESS KEY", base url=" https://n.tokeness.io/v1 https://n.tokeness.io/v1 ", resp = client.chat.completions.create model="deepseek-v4-flash", copied exactly from their model marketplace messages= {"role": "user", "content": "hello"} , The good One key for everything. New models land fast — kimi-k3, grok-4.5, the gpt-5.6 series, gemini-3.6-flash are all live as of late July. Pricing transparency is better than most resellers: every model shows input/output price per 1M tokens in the console. No "30% of official, do the math yourself" nonsense. deepseek-v4-flash is updated to the official 0731 release and billed 40% below official pricing. It's been solid for agent workloads in my testing. Free tier that actually works: tokeness/free routes to ~200B-param models DeepSeek V4 Flash class . Enough to test your whole integration without paying. Business invoices + per-request usage logs status, model, cost per call . If you expense API costs, this matters. They run an impact program — 10M to 5B free tokens/month for OSS/nonprofits, with public rules docs link . The bad GPT, Claude and some others don't work from mainland China IPs. They announced this June 15 as a compliance move. Overseas networks are fine; Chinese models work from anywhere. Weirdly, publishing this restriction makes me trust them more than resellers promising "direct Claude access from China," but it's a hard limitation depending on where you are. It's still a third-party relay. They claim they only forward and retain nothing — you can't audit that. My rule: personal projects fine, customer data/prod secrets go through official APIs only. Same rule I'd apply to any reseller. Prices aren't locked. They state upfront that pricing follows upstream costs, FX and channel policy. gpt-5.6-luna dropped 80% recently, but cuts can presumably reverse. "Original models, no dilution" is their claim. I ran basic sanity checks identity questions, known-answer reasoning prompts, response field structure and nothing looked off, but YMMV. Who it's for Multi-model users who want one bill and proper invoicing, and have the network situation for GPT/Claude sorted. Not for anyone whose core need is "Claude directly reachable from mainland China" — they're explicit that this isn't that. They also have a referral program 20% of a friend's first top-up — mentioning for completeness, deliberately not posting my link. Anyone else using it? Curious how their stability holds under heavier agent workloads.