PARAS MADAN
All writing

Latest in AI

Tencent Hy4 Flash API: Pricing, Discount Dates and Limits

The cheaper Hy4 model is available now, but provider pricing displays and promotion deadlines need separate checks.

Tencent released Hy4 Flash on November 16, with list prices of about $0.070 per million input tokens and $0.281 per million output tokens. Vercel also added it to AI Gateway. For developers considering a cheaper model, the immediate questions are which price applies and whether the workflow still passes its checks.

Hy4 Flash pricing: list rates versus the discount

Tencent's launch announcement gives these approximate US-dollar list rates and a 50% TokenHub promotion through November 30. The discounted column below is arithmetic on those published rates, not a captured invoice.

Token categoryTencent list price per millionCalculated at 50% of list
Uncached input$0.070$0.035
Cached input$0.003$0.0015
Output$0.281$0.1405

Vercel's launch notice specifies its discount on input, cached input and output through November 29, 2026. Do not substitute Tencent's November 30 wording when budgeting the Vercel route. Neither announcement establishes a shared cutoff hour or timezone for these two offers.

The AI Gateway model catalogue displays discounted prices rounded to $0.04 input and $0.14 output per million tokens, alongside $0.0015 cached input. Those displayed figures should not be treated as a contradiction of the more precise arithmetic above, or as proof of the exact precision used on an invoice. Check the selected provider's billing view before committing a large workload.

What a document batch would cost

Suppose a proposed batch uses ten million uncached input tokens and two million output tokens. At Tencent's published list rates, the token subtotal is:

Input: 10 × $0.070 = $0.700
Output: 2 × $0.281 = $0.562
Total: $1.262
At 50% of those rates: $0.631

This is a worked estimate, not a benchmark or an observed bill. It assumes those token counts already include every model call and excludes any other charges. It does not assume a cache hit merely because documents repeat.

For a document-processing application, the more useful comparison is the cost of accepted outputs. If a cheaper run produces a malformed table that needs another attempt, count both attempts. If someone must repair the result manually, track that time separately instead of hiding it inside a low token subtotal.

Calling Hy4 Flash through AI Gateway

Vercel's model identifier is tencent/hy4-flash. Its announcement documents this minimal AI SDK pattern; the prompt below is an illustrative replacement, not a test I ran:

import { streamText } from 'ai';

const result = streamText({
  model: 'tencent/hy4-flash',
  prompt: 'List three checks for a generated invoice table.',
});

for await (const text of result.textStream) {
  process.stdout.write(text);
}

The catalogue's setup guidance uses an AI Gateway key in the AI_GATEWAY_API_KEY environment variable. Keep that credential on the server, outside committed source code. A model name alone does not configure authentication or establish permission to send a document to the provider.

For coding agents, Vercel's launch instructions use npx vercel ai-gateway setup, followed by selecting the model in the agent. Treat that as configuration work: review the changes to your existing agent settings rather than assuming it leaves every provider preference untouched.

Context capacity is not output capacity

The Gateway catalogue lists a 1M-token context window and a 64K maximum output. Its API reference says prompt and response share the context window and reasoning tokens count toward the output cap. Set a smaller output budget appropriate to the task rather than interpreting a million-token context as permission to generate a million-token answer.

Vercel lists image input, tool calling, structured outputs, prompt caching and preserved reasoning across turns. Those capabilities make Flash worth evaluating for repeated document and agent work, but a supported feature is not an accuracy guarantee.

A useful first comparison is one bounded job: extract a table, preserve its units and flag missing values. Test a clean document, a scanned page and an ambiguous row against the same acceptance criteria. Record validation failures, retries, latency and total usage. These are proposed checks, not results from this article.

When to keep the larger Hy4 model

Tencent recommends Flash for everyday work and the larger Hy4 for difficult, long-running software engineering. It also plans to open-source Flash; the launch announcement is not proof that downloadable weights are already available.

I would avoid changing an entire agent workflow on the strength of a promotional token price. Start where a result can be checked cheaply, then expand if the completed work justifies it. The durable saving is a lower cost per accepted result after the promotion ends, not the smallest number on launch day.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams turn model launches, pricing changes and technical workflows into clear content that developers can use.

Book a 30-minute call