Tencent Hy4 Pricing: Reasoning Changes the Bill

Tencent released the full version of Hy4 on 4 November 2026, and on OpenRouter it lists at $0.417 per million input tokens and $1.25 per million output tokens. That is half of the Hy4 preview’s rate, and it looks like a mid-priced model with a very large context window. For a support chatbot, that rate card hides the number that matters: how much the model thinks before it answers.

On the day it launched, we sent the same support-bot request from our own WordPress server to Hy4 with its reasoning at default, set to low, and switched off, then to the Hy4 preview, Hy3 and five rivals. We timed every answer, recorded the tokens each one billed, and checked each answer against our documentation.

The short answer: with default settings, Hy4 cost $20.45 per 1,000 answers and took a median of 162 seconds to finish each one, because it wrote about 14,900 reasoning tokens per answer before a reply of about 120. Setting effort to low changed nothing ($20.93, 164 seconds). Switching reasoning off cut it to $1.80 per 1,000 answers with no cache, $0.21 with a warm cache, and 2.26 seconds, and the answers were just as accurate. Claude Haiku 5.5 cost $0.89 on the same request. If you run MxChat, read the section on the 60-second limit before you pick this model.

Tencent Hy4 pricing at a glance

These are the rates on OpenRouter’s model API on 4 November 2026. Tencent is the only host listed for the final Hy4 so far.

The discount has an end date on Tencent’s side. Tencent Cloud’s TokenHub price page lists Hy4 at 6 yuan input, 18 yuan output and 0.3 yuan for a cache hit per million tokens, and says calls are billed at 50% of that until 23:59:59 Beijing time on 31 December 2026. OpenRouter’s listing carries a 50% discount flag but gives no end date, so check it before you budget for next year. Doubling the rates doubles every Hy4 cost in this article.

OpenRouter lists text and image input, and up to 65,536 output tokens on Tencent’s endpoint. Tencent says it plans to release the final model’s weights. We could find weights only for the August preview, not for this version, which is why there is only one host today.

How we tested it

Every request was the one a MxChat support bot actually sends: the site’s real system prompt plus four passages pulled from our own documentation, about 4,150 tokens on Hy4’s count, followed by one visitor question. We asked four questions, each twice:

1. Does MxChat work with WooCommerce, and can the bot look up an order status for a logged-in customer?
2. Do I need Pinecone to use the knowledge base, or can I store embeddings in WordPress?
3. Can a visitor be handed over to a human on Slack or Telegram when the bot cannot answer?
4. Can I use my own OpenRouter or OpenAI API key, and which models can I choose from?

We ran eleven setups through OpenRouter, with no fallbacks on the pinned ones: Hy4 pinned to Tencent at default, at effort low and with reasoning off; the Hy4 preview on Tencent and on DeepInfra; Hy3 on Tencent; Mercury 3 on Inception; and Claude Haiku 5.5, DeepSeek V4.1 Flash, GPT-5.6 Luna and MiniMax M3.1 Flash on whichever host OpenRouter picked. That is 88 requests.

We report cost two ways. No cache is the tokens each answer used multiplied by the list price, which is close to what a low-traffic site pays. Warm is what OpenRouter actually billed on the second pass, after the first pass had filled the prompt cache. For DeepSeek V4.1 Flash, which OpenRouter routes across several hosts, we use the billed first-pass cost as the no-cache figure.

What 1,000 chatbot answers cost

Reasoning is about 91% of the default bill. Input cost $1.73 per 1,000 answers. The visible reply cost about $0.15. The other $18.58 was reasoning tokens, billed at the output rate. One answer alone used 23,650 of them, to reply to a question about API keys.

A cache can’t rescue it. Hy4 cached well, with about 4,048 of 4,150 input tokens read from cache. But input was already the small part of the bill, so the warm default figure only fell to $16.03. Without reasoning, the same cache takes the bill from $1.80 to $0.21, a drop of 88%.

The cheaper model costs more than the dearer one. The Hy4 preview has twice the rate on every line and came in at $11.44 per 1,000 answers against $20.45, because it wrote about 3,100 reasoning tokens where the final Hy4 wrote about 14,900. For this kind of request, the final version thinks about five times as long as the preview did.

A site handling 10,000 questions a month would pay about $205 at Hy4’s default with no cache, about $18 with reasoning off and no cache, and about $2 with reasoning off and a warm cache. When the discount ends those figures double, to about $410, $36 and $4.

Effort low doesn’t lower the effort

This was the surprise of the test. OpenRouter lists a reasoning-effort setting for Hy4, and we sent "reasoning": {"effort": "low"}. The request was accepted, and the model spent 15,230 reasoning tokens per answer, slightly more than at default. Eight answers each isn’t enough to say low is worse, since single answers ranged from about 5,000 to 24,000 tokens at both settings. It is enough to say low didn’t help.

The setting that works is the off switch: "reasoning": {"enabled": false}. Every one of the eight requests came back with zero reasoning tokens and took between 2.12 and 2.52 seconds. Some recent models refuse this outright, including StepFun Step 5 and Liquid AI’s LFM 2.5, which return HTTP 400 when you try. Hy4 accepts it.

We tested through OpenRouter on launch day. Tencent may tune the model’s default behaviour, and TokenHub’s direct API may behave differently, so test your own traffic before you rely on these numbers.

Speed

At default, the fastest of the eight Hy4 answers took 50.55 seconds and the slowest took 230.26. Seven of the eight took more than a minute. No visitor to a support widget waits three minutes for a reply, and this is the bigger problem at default settings, even ahead of the cost.

With reasoning off, Hy4 is in the same range as the fast tier at 2.26 seconds: a little quicker than Haiku 5.5 (2.37 s) and slower than GPT-5.6 Luna (1.51 s), DeepSeek V4.1 Flash (1.06 s) and Mercury 3 (0.68 s).

What happens if you pick Hy4 in MxChat today

We’d rather you read this here than find it on your live site. MxChat’s OpenRouter request sends the model, the messages, a temperature and a streaming flag. It doesn’t send a reasoning setting, so Hy4 runs at its default. MxChat also has a 60-second limit on that request.

So we reproduced that exact request from the same server, four times, one for each question, with streaming on and a 60-second cut-off. All four hit the limit with no answer text at all. Reasoning started streaming after about 1.5 seconds, but the model was still thinking when the 60 seconds ran out. This was a replay of the request, not a run through the live widget, but the outcome is plain from the numbers above: a visitor on Hy4 would generally not get an answer.

Our advice for MxChat sites is to leave Hy4 alone for now and use one of the models in the next table. MxChat has no setting to turn reasoning off on OpenRouter models. If you call Hy4 from your own code, send "reasoning": {"enabled": false} on every request.

Is Hy4 any better with the thinking on?

Not on this job. On every setup, all the answers contained the core facts from our documentation: order lookups need the Pro add-on and a logged-in customer, Pinecone is optional, Slack and Telegram handoff both exist, and you can bring your own API key.

The difference was in how they read. With reasoning at default or low, 3 of 8 answers added that the bot didn’t have enough information about part of the question, and many ended with a sales pitch for a license upgrade. With reasoning off, 1 of 8 hedged that way, and the answers averaged 37 words against 72. Here is Hy4 with reasoning off on the WooCommerce question, word for word:

Yes. The WooCommerce add-on requires an active MxChat Pro license and supports AI order inquiries, recent orders with status, and direct order links. Order history access is for logged-in users only.

That is what a support bot should say. Eight answers per setup is a small sample, and reasoning may well earn its cost on the long agent work Tencent built Hy4 for. A four-passage support question simply isn’t that job, and the 15,000 extra tokens didn’t produce a better reply here.

How Hy4 with reasoning off compares

Warm, reasoning-off Hy4 is competitive at $0.21 per 1,000 answers, ahead of Haiku 5.5, which reported no cached tokens on this request. Cold, it is about twice the price of Haiku 5.5 and nearly four times the price of Mercury 3. So it only makes sense on a site with enough traffic to keep the cache warm, and that margin narrows further when the discount ends: $0.42 warm and $3.61 cold.

Two of these figures depend on routing. DeepSeek V4.1 Flash and MiniMax M3.1 Flash went to hosts OpenRouter chose, Relace and Wafer for the first and Tenstorrent for the second, so your host and bill may differ. MiniMax’s $0.82 is at its listed rate; Tenstorrent actually billed $0.24 on the first pass. Mercury 3’s figures are also at a launch discount, $0.10 input and $0.35 output, with no end date listed.

Two reliability notes from launch day

All 24 requests to the final Hy4 came back with HTTP 200, although 3 needed a second attempt. The preview was rougher: the Tencent-pinned setup lost one request after four attempts to HTTP 502, which is why that row covers 7 answers, and another needed three. Pinned to DeepInfra, the preview returned HTTP 429 “temporarily rate-limited upstream” on seven questions and timed out on the eighth, even with retries, so we have no DeepInfra figures. It is a launch day, so treat this as a snapshot.

Mercury 3, which returned HTTP 429 on every request when we tried it on 2 November, answered all eight today on Inception at the first attempt.

How to choose a model in MxChat

1. Create an API key on OpenRouter and add credit to the account.
2. In WordPress, open MxChat’s settings and paste the key into the OpenRouter field. Our documentation covers where each key goes.
3. Choose a model that answers without long reasoning, such as inception/mercury-3 or anthropic/claude-haiku-5.5.
4. Send your bot a few real visitor questions and check the answers against your own pages before you switch it on for everyone.
5. Check OpenRouter’s usage for the first few days and look at output tokens per request. A model writing thousands of output tokens for a short reply is spending them on reasoning.

If you’re deciding between OpenRouter and going direct, our OpenRouter vs direct API pricing guide covers what the router adds. Order lookups for logged-in WooCommerce customers need MxChat Pro. Reasoning slowed another big model too: our Kimi K3.1 Pro test measured a 16-second median at default and under 7 seconds with reasoning off. And DeepSeek V4.1 Flash pricing covers the cheapest option in this test.

Frequently asked questions

How much does Tencent Hy4 cost?

On OpenRouter on 4 November 2026, Hy4 costs $0.417 per million input tokens, $0.021 per million cached input tokens and $1.25 per million output tokens, which is a 50% launch discount on $0.834, $0.042 and $2.50. Tencent Cloud TokenHub lists 6, 0.3 and 18 yuan per million, billed at 50% until 31 December 2026.

What does Hy4 cost per chatbot answer?

On our support-bot request, $20.45 per 1,000 answers at default reasoning, $1.80 with reasoning off and no cache, and $0.21 with reasoning off and a warm cache. Those figures double at full rate.

Why is Hy4 so slow?

At default it writes about 14,900 reasoning tokens before a short answer, at roughly 100 tokens per second. That gave a median of 162 seconds. With reasoning off, the median was 2.26 seconds.

How do I turn off reasoning on Hy4?

On OpenRouter, send "reasoning": {"enabled": false} in the request. Setting effort to low didn’t reduce reasoning in our test.

Can I use Hy4 in MxChat?

We don’t recommend it today. MxChat doesn’t send a reasoning setting to OpenRouter and limits the request to 60 seconds. All four of our replays of that request timed out with no answer text.

Is Hy4 open source?

The August preview’s weights are public. Tencent says it plans to release the final Hy4’s weights, and we couldn’t find them on launch day. Tencent is the only host on OpenRouter for the final version so far.

Is Hy4 the same as Hunyuan?

Hy is Tencent’s current name for its Hunyuan model family. OpenRouter also lists older models from the line, such as tencent/hy3 and tencent/hunyuan-a13b-instruct.

Method note: tests run 4 November 2026 from mxchat.ai’s server in the US through OpenRouter, 4 questions × 2 passes per setup, temperature 1. Prices are OpenRouter rates read the same day. Costs per 1,000 answers are means over the answers in each setup. The three Hy4 setups ran side by side and can share a cache, so first-pass billing is not a clean cold figure; the no-cache column uses token counts and list rates to avoid that. Your costs depend on your prompt size, answer length and traffic pattern.