Tencent's Hy4 Flash joins Vercel's AI Gateway at one-tenth the cost of the full Hy4

- Reading
- 4 min
- Tables
- 1
- Code
- 2
Vercel added Tencent's Hy4 Flash to the AI Gateway with an introductory 50% discount through November 29, 2026. The model keeps the full Hy4's features, such as a 1-million-token context window and reasoning that can be turned off, at about one-tenth the price per token.
A Chinese model priced at one-tenth
Vercel confirmed, in its official changelog, that Hy4 Flash, from China's Tencent, is now available on the AI Gateway with an introductory 50% discount on input, output, and cache tokens, valid through November 29, 2026. Hy4 Flash is the cheapest model in the Hy4 family and, according to Vercel, delivers the full set of Hy4 features for about one-tenth the price per token charged by the family's flagship model.
There are two stacked savings here, not just one. The structural one, around 90% compared to the full Hy4, is permanent and part of the Flash's positioning. The promotional one, an additional 50%, only lasts until the end of November: after that, the price returns to the Flash's normal level, which is already low, but without the extra cut.
What Hy4 Flash keeps (and what it cuts)
For anyone evaluating a model by what it delivers, this isn't a capped-down version of Hy4. The list of preserved features includes reasoning with "thinking" memory carried across conversation turns, tool calling, structured output, prompt caching, image input, and a 1-million-token context window, the same as the full model.
The cut comes elsewhere: you can turn off reasoning on requests that don't need it. Vercel itself recommends this for high-volume workloads, such as code generation, agents, and document processing, where paying for step-by-step reasoning on every call is wasteful.
In short: Hy4 Flash isn't a model with fewer features, it's the same feature set as Hy4 with a switch to turn off reasoning when the use case doesn't justify that cost.
Getting the model up and running
On the AI Gateway, the model's name is tencent/hy4-flash. With the AI SDK, the call follows the same pattern as any other model on the gateway:
import { streamText } from 'ai';
const result = streamText({
model: 'tencent/hy4-flash',
prompt: 'Summarize the tradeoffs of server-side rendering in three sentences.',
});
for await (const text of result.textStream) {
process.stdout.write(text);
}To use the model inside code agents like Claude Code, Codex, or Cursor, the way to go is to install the latest version of the Vercel CLI and run the setup wizard:
npx vercel ai-gateway setupThen just select tencent/hy4-flash from the agent's model list. Anyone who wants to test before switching the production model can use the AI Gateway's own model playground.
Where this intersects with front-end work
In the day-to-day work of building interfaces, Hy4 Flash fits into three areas that already run on top of the AI Gateway: generating components from structured output, extracting data from documents or screenshots (image input), and high-volume autocomplete inside code agents. None of these tasks require, in most calls, the full step-by-step reasoning.
| Scenario | Reasoning on | Reasoning off |
|---|---|---|
| Generate a component from a short prompt | Unnecessary in most cases | Cheaper and faster |
| Plan a refactor across multiple files | Recommended | Risk of a shallow response |
| Extract structured data from a document | Optional | Sufficient in most cases |
| Debug a complex error in a code agent | Recommended | Risk of an incomplete solution |
This breakdown doesn't come ready-made from Vercel: it's the reading that makes sense given what the changelog describes about the reasoning toggle. It's worth validating case by case before locking it in as a production default.
Why this matters for anyone with an agent already in production
Swapping the model behind an agent is, in theory, trivial: change the model string and you're done. In practice, what really changes is the cost calculation per task, because now you can keep reasoning on only for calls that require step-by-step thinking and turn it off for the rest, within the same pricing contract and the same model.
The AI Gateway helps measure this effect without switching tools. The usage and cost report built into the platform, routing rules, and per-API-key budgets give visibility into how much each route, with reasoning on or off, is costing before it becomes a fixed default in the code.
What's still left open
Vercel guarantees that the AI Gateway reflects the provider's price with no markup and no platform fee on inference, including on requests with a bring-your-own key (BYOK). That clears up the question of hidden fees, but it doesn't answer the more annoying question: how good is Hy4 Flash's reasoning really, compared to similarly priced Western models, such as the leaner versions of GPT or Claude.
The changelog doesn't include a comparative benchmark, just the feature description and the price relationship to Hy4 itself. Before moving production workload to Flash, it's worth running your own set of tasks (component generation, structured extraction, tool calling) and comparing cost and quality side by side in the model playground, rather than trusting the discount alone.
Another point that weighs heavily for anyone handling sensitive data is the model's origin: being a Chinese provider, data retention evaluation matters more here than with domestic models. The AI Gateway already supports Zero Data Retention, but it's worth confirming whether that setting is specifically available for tencent/hy4-flash before running prompts with customer data.
Translated from the Brazilian Portuguese original · Read the original
Vercel Sandbox now accepts managed npm credentials for AI agents to install private packages
Vercel Sandbox integrates the team's npm credentials, stored as Shared Environment Variables, so a code agent can install private dependencies without ever seeing the access token.





