Yesterday I wrote about Tencent's new Hy4 Flash model and its almost silly price: 0.07 dollars per million input tokens. A few readers asked the obvious follow-up question. Fine, it is cheap, but how do I actually use it without rebuilding my whole setup? The good news is that Tencent designed this model to be a drop-in guest in other people's software. Hy4 Flash officially supports the OpenAI Chat Completions protocol, the OpenAI Responses protocol and the Anthropic Messages protocol, which means most tools that accept a custom endpoint can talk to it today. Here is the full setup, the way I would do it, in about ten minutes.
What you need before you start
- A Tencent Cloud account. Registration is free; you only pay for what you use.
- A payment method the platform accepts for topping up a small balance. Start with the smallest amount you can. Five dollars at these prices is a lot of tokens.
- One app you already use that allows a custom API endpoint or a custom OpenAI-compatible provider. Most serious writing, coding and automation tools have this setting, often labelled custom endpoint, base URL or provider.
- Ten quiet minutes and a small test prompt you can judge easily.
Step 1: Find the model and read its card
Sign in to Tencent Cloud and open the language model section of the console, the area the documentation links to TokenHub. Find hy4-flash in the model list. Before touching any keys, read its card properly: one million token context window, up to 960,000 tokens of input, up to 64,000 tokens of output, image understanding, and tool calling. Note the three prices as well: 0.07 dollars per million input tokens, 0.281 dollars per million output tokens, and 0.003 dollars per million cached input tokens. Knowing the cached price matters later, because it rewards apps that resend the same long instructions.
Step 2: Create your API key and guard it
In the console, create an API key for the model service. Copy it once and store it in a password manager or a private note that only you can open. Treat this key like a bank card number. Anyone who has it can spend your balance, and keys pasted into public code or screenshots are one of the most common ways small creators lose money to strangers. If a key ever leaks, delete it and make a new one immediately.
Step 3: Point your existing tool at Tencent
Open the settings of the app you want to connect and find its custom endpoint or provider section. You will make three changes. First, paste the API endpoint address exactly as shown in Tencent Cloud's API guide for the language model service. Copy it from the guide rather than typing it from memory, because one wrong character means a dead connection. Second, set the model name to hy4-flash, exactly as it appears in the model list. Third, paste your API key. Save the settings. That is the whole trick: because Hy4 Flash speaks the OpenAI and Anthropic protocols, your app believes it is talking to a familiar model, while your requests quietly go to Tencent's servers instead.
Step 4: Test small, then check the bill
Run one small, real task first: summarise a page you have already read, rewrite a paragraph in your own voice, or explain a short piece of code you understand. Judge the answer yourself. Then open the usage or billing page in the console and look at what that test cost. At these prices it should be a tiny fraction of a cent, and seeing that number with your own eyes is the moment the price war stops being a headline and starts being your workflow. Only after the test passes should you route real client work through it.
Step 5: Set a spending limit before you scale up
Automation has a way of quietly looping, retrying and multiplying costs while you sleep. Before you connect any agent or bulk workflow, set the tightest budget alert or balance limit the console offers, and keep your topped-up balance small on purpose. A five-dollar balance can do a month of routine work at Flash prices, and it also caps your worst day if a script misbehaves.
How I would split the work in real life
- Send Hy4 Flash the routine 90 percent: first drafts, summaries, translations, formatting, simple code, captions and variations.
- Keep your strongest model for the work that wins or loses clients: strategy, tricky reasoning, final polish and anything confidential.
- Turn on prompt caching where your app supports it, especially if you resend a long instruction file or style guide with every request. At 0.003 dollars per million cached tokens, that habit nearly deletes your input bill.
- Never send a client's private documents, contracts or unreleased material to any hosted API, from any country, until you have read its data terms and your client is comfortable.
The honest bottom line
None of this requires new hardware, a new subscription or learning a new app. That is exactly why this release matters to small creators and freelancers like us. The expensive part of AI was never the interface. It was the meter running behind it, and Tencent just slowed that meter to a crawl. Set it up once, test it honestly, cap your spending, and let the cheap model carry the heavy routine work while you save your money and your best tools for the jobs that actually need them.
Top comments (0)