Why I Built It
Over the past year Claude Code and Codex have become my daily drivers, and with that came a new headache: wrangling API endpoints. One entry point for Claude, another for OpenAI, yet another for Gemini — each with its own pricing rules, and cache pricing I had to go look up separately every time. Rather than keep patching together accounts from every direction, I decided to run my own relay and pull the models I use most under a single endpoint. That’s how this site came to be: MIHOUO, at https://api.mihouo.com.
It runs on a BandwagonHost VPS, powered by the open-source new-api. Registration is open — sign up, top up your balance in the dashboard, and you’re good to go.
Models and Pricing
Prices are in USD per million tokens, pay-as-you-go. Every series is discounted off the official API price (GPT as low as 7% of list, GLM at 15%, Claude starting at 20%). In the tables below, strikethrough is the official price and bold is ours.
Claude Series
| Model | Input | Output | Discount |
|---|---|---|---|
| claude-opus-5-5 | 20% of list | ||
| claude-sonnet-5-5 | 20% of list | ||
| claude-fable-5-2 | 30% of list | ||
| claude-haiku-5-5 | 95% of list |
Above a 100k context, haiku’s official rates go up to $0.5 input / $2.5 output; ours are $0.475 / $2.38.
GPT Series
| Model | Input | Output | Discount |
|---|---|---|---|
| gpt-6-astra | 7% of list | ||
| gpt-6.1-sol | 10% of list | ||
| gpt-6-luna | 95% of list |
Past 272k context, official prices for astra and sol go up; the discount stays the same: astra at $1.4 input / $5.25 output, sol at $0.4 / $1.5. The image model gpt-image-2.6 is billed per image at $0.013.
Gemini Series
| Model | Input | Output | Discount |
|---|---|---|---|
| gemini-4-argon | 40% of list | ||
| gemini-3.8-flash | 25% of list |
Chinese Models
| Model | Input | Output | Discount |
|---|---|---|---|
| glm-5.3 | 15% of list | ||
| glm-5.3-flash | 50% of list | ||
| deepseek-v4-pro | 65% of list | ||
| deepseek-flash | 65% of list | ||
| kimi-k3.1-pro | 45% of list |
The two DeepSeek models use time-based pricing: 9:00–12:00 and 14:00–18:00 Beijing time are peak hours with roughly 25% higher rates (deepseek-v4-pro peaks at $0.858 input / $2.574 output on our site). Off-peak runs at the lower prices in the table — scheduling batch jobs at night is noticeably cheaper.
Cache reads and writes are discounted too — sonnet-5-5’s official cache read is $0.1, ours is $0.02. With cache-heavy tools like Claude Code, most of the context is served from cache, so the real bill comes in well below what the headline prices suggest.
Running the Numbers on My Own Last 30 Days
Calling something “cheap” is abstract, so let’s price out my actual usage. The chart below is from cc-switch on my machine, covering the last 30 days: 31,092 requests, 99.2M input tokens, 22.8M output, 4.65M cache writes, and 5.57B cache hits — a 98.2% cache hit rate. The classic profile of a heavy Claude Code user.

Now imagine that entire 30-day workload running on a single model, and price it at official rates versus ours:
| Model | Official | Ours | Discount |
|---|---|---|---|
| gpt-6-astra | $7,760 | $543 | 7% of list |
| claude-fable-5-2 | $2,866 | $860 | 30% of list |
| claude-opus-5-5 | $1,990 | $398 | 20% of list |
| gemini-4-argon | $1,967 | $787 | 40% of list |
| glm-5.3 | $1,687 | $253 | 15% of list |
| kimi-k3.1-pro | $1,540 | $693 | 45% of list |
| gemini-3.8-flash | $1,155 | $289 | 25% of list |
| claude-sonnet-5-5 | $995 | $199 | 20% of list |
| gpt-6.1-sol | $995 | $100 | 10% of list |
| deepseek-v4-pro | $466 | $242 | 65% of list |
| glm-5.3-flash | $193 | $97 | 50% of list |
| deepseek-flash | $91 | $29 | 65% of list |
| claude-haiku-5-5 | $78 | $74 | 95% of list |
| gpt-6-luna | $78 | $74 | 95% of list |
A few notes:
- The same workload costs anywhere from tens to over seven thousand dollars at official prices. On this site, the heavily discounted models all come in at 20% of list or less — gpt-6-astra drops from $7,760 officially to $543 here, a straight 93% saving.
- The $642.84 in the chart is what my actual mixed usage would cost at official unit prices; run through this site instead, the same mix comes to roughly $130.
- The official column for the two DeepSeek rows shows peak-hour pricing while ours shows off-peak — even priced entirely at peak hours, ours would only be $303 / $59.
- Cache hits dominate my usage (5.57B — 56× the fresh input), so long sessions live on cache and the cache-read rate effectively decides the total bill. Ours discounts cache reads too, which matters more than anything else for Claude Code users.
Getting Started
The site speaks both the OpenAI and Anthropic protocols, and most models work over either endpoint, so common clients can hook right in.
Generic OpenAI Clients
Just point the Base URL at the site:
Base URL: https://api.mihouo.com/v1API Key: sk-xxxx (create one under the "Tokens" page in the dashboard)Claude Code
Claude Code uses the Anthropic protocol — two environment variables and you’re connected directly:
export ANTHROPIC_BASE_URL=https://api.mihouo.comexport ANTHROPIC_AUTH_TOKEN=sk-xxxxOther Anthropic-Protocol Clients
The same goes for Codex, Cursor, OpenCode, and any other tool that lets you set a custom Anthropic endpoint: point the base URL at https://api.mihouo.com and use a site token as the key.
Fine Print
- Prices track upstream costs; the “Model Pricing” page in the dashboard always has the current numbers.
- Discounts are a straight multiplier on official API prices — no minimum spend, no monthly fee, your balance simply runs out.
- This site is primarily for my own use, opened up for others along the way. There’s no support team — if you hit a problem, drop a comment and I’ll get to it.
If juggling multiple endpoints and accounts is wearing you down, or you just want a cheaper Anthropic endpoint for Claude Code, come grab an account at MIHOUO and see how it feels.