Hy3 API: toman pricing and code

What is Hy3 good for?

Hy3 is one of Tencent's models. Put "tencent/hy3" in the model field and the rest of your code stays as it is. Hy3 keeps 262,144 tokens in view at once, and that number is what caps the conversation history in your product. A single response can run to 128,000 tokens. On price it sits in the "cheap" band — cheaper than 298 and dearer than 100 of the other paid models in the catalogue.

What it can do beyond plain text: it supports tool calling, so it can invoke your own functions with valid arguments; it returns schema-valid JSON through response_format, ready to hand to your code; it has a reasoning mode that pays off on multi-step problems, maths and debugging. All of it works through the standard parameters of the official OpenAI SDK — no custom client, no wrapper. Keep in mind that reasoning tokens are output tokens and do appear on the bill.

Pricing is 45,701 toman per 1M input tokens and 182,804 per 1M output tokens. A 1,000-word round trip on Hy3 lands near 297 toman. It supports cached input: repeated context is billed at 11,425 toman per 1M, which matters a lot if your system prompt is long. You are always charged for the usage the request actually reported, never for the estimate, and a failed request costs nothing.

The closest alternative with the same capabilities from a different provider is Llama 3.3 70B Instruct: Hy3 works out about 9% cheaper, and its context window is larger. Both run on the same key and the same code, so trying the other one is a single string change.

To make the figure concrete: 100,000 toman of credit buys roughly 337 thousand-word requests on Hy3, and every 1,000 toman is about 3,367 words of round trip. A job with one million input tokens and one million output tokens comes to 228,505 toman in total. Filling this model's 262,144-token window costs 11,980 toman on the input side alone, which is the real reason to keep conversation history short.

Its nearest relative in the catalogue is "tencent/hy4-preview", and choosing between those two is where most people hesitate. On a thousand-word request this variant works out about 5.1× cheaper (297 against 1,501 toman). The context windows differ too: 262,144 against 1,048,576 tokens. Maximum answer length differs as well: 128,000 against 64,000 tokens. On parameters the other takes logprobs, parallel_tool_calls, top_logprobs.

Tencent has 8 models in our catalogue; the cheapest is Hy-MT2-1.8B at 99 toman per thousand words and the dearest Hy4 preview at 1,501. Among the less common parameters it accepts logit_bias, max_completion_tokens, min_p, presence_penalty, reasoning_effort, repetition_penalty — all through the standard request body. By context size the nearest option from another provider is Aion 3.5 at 262,144 tokens.

Good fits: high-volume work such as classification, tagging and bulk summarising, logic puzzles and code review, agents that reach out to APIs and databases, extracting data against a fixed schema, analysing a long document or codebase in one request.