Vercel announced Hy4 Flash on AI Gateway on November 16, 2026, under the model ID tencent/hy4-flash. Its 50% introductory discount on input, cached-input, and output tokens runs through November 29; the catalog lists separate rates for each category.
What does Vercel charge for Hy4 Flash?
Vercel’s catalog displays these promotional rates in U.S. dollars per million tokens:
| Billing category | Vercel-displayed rate | Offer term |
| Input | $0.04 | 50% discount through November 29, 2026 |
| Output | $0.14 | 50% discount through November 29, 2026 |
| Cached input (read) | $0.0015 | 50% discount through November 29, 2026 |
Vercel displays the input and output prices to two decimal places. It says AI Gateway reflects provider pricing without markup and charges no platform inference fee, including for requests using BYOK—bringing your own key. The displayed rates and billing details appear on Vercel’s Hy4 Flash model page; the discount announcement gives its scope and end date.
Vercel also says free users who have not made a payment receive $5 in AI Gateway credits every 30 days. Hy4 Flash usage is billed at API rates, so that general credit offer is not a model-specific free tier.
How do you call the model through AI Gateway?
Use tencent/hy4-flash as the model ID. Vercel documents calls through the AI SDK, OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats. Its setup instructions use an AI Gateway API key stored in the AI_GATEWAY_API_KEY environment variable; for the API formats, Vercel says to change the base URL.
The listing includes text and image input, reasoning, tool calling, structured outputs, and prompt caching. Images count toward input-token usage and the context window. Vercel lists a 1-million-token context and a maximum output of 64,000 tokens, with reasoning tokens included in that output limit. These interface options and limits are detailed in Vercel’s model documentation.
When should you use only or order for provider routing?
Use only when requests must stay within a specified provider allowlist. It is a hard restriction: if none of the allowed providers is available, the request fails. Use order to set provider preferences while leaving unlisted providers eligible as fallbacks.
If a request should continue through another eligible provider when your preferred one is unavailable, useorder;onlyintentionally makes the allowed-provider list a hard boundary.