RSS
Research

Tencent Hy4 flash: price, coding benchmarks and how it compares to Claude Haiku 5.5

Tencent's open-weight Hy4 flash costs $0.070 input and $0.281 output per million tokens, and OpenRouter currently charges half that. Tencent's own tests put it close to Claude Haiku 5.5 on office tasks but far behind DeepSeek V4.1 Flash on terminal coding.

Rack of electronic equipment in a dark room illustrating Tencent Hy4 flash pricing and benchmarks
Photo by Tyler on Unsplash

Tencent's Hy team released Hy4 flash on November 16 as an open-weight model: 12B active parameters out of roughly 220B, text and image input, and a 1 million token context window. Its list API price is $0.070 per million input tokens and $0.281 per million output tokens, and it is now live on OpenRouter as tencent/hy4-flash, currently at half the list price.

The pitch is cost. The more useful detail is in Tencent's own benchmark table: Hy4 flash is the cheapest model on every row, but on the terminal-style coding tests it trails DeepSeek V4.1 Flash by a wide margin, and Tencent itself says to use the larger Hy4 for long-horizon software engineering. Below are the published numbers, the price math against Claude Haiku 5.5, and how to try it. We have not run our own test; every number names its source and was read on November 18, 2026.

What is Hy4 flash?

  • Open weights: Hy4 flash (instruct), Hy4 flash-FP8 and Hy4 flash-base on Hugging Face, ModelScope and GitCode.
  • Size: 12B active parameters per token (11.3B text plus 0.43B vision). The Hugging Face card says 220B total including 64B of n-gram embedding (Engram) parameters (63.5B in its detailed table); Tencent's announcement page says 224B. Both say 12B active.
  • Context: 1,048,576 tokens. OpenRouter lists a 65,536-token output cap.
  • Architecture: sparse and sliding-window attention interleaved 1:3, mixture-of-experts routing in a smaller latent space (16 of 512 experts per token), and the Engram bigram and trigram lookup. Tencent says this cuts the KV cache to 1/6 of Hy4's.
  • Reasoning: reasoning_effort accepts "high" (the default, thinking on) or "no_think". Recommended sampling is temperature 0.9, top_p 1.0.
  • Licence: Tencent Hy License 1.0, derived from Apache 2.0 with extra commercial terms for model-as-a-service businesses (details below).

How much does Hy4 flash cost?

Tencent's announcement lists these prices per million tokens. Through November 30, it offers 50% off API pricing on Tencent Cloud TokenHub and free use inside its WorkBuddy app.

  • Hy4 flash: $0.070 input, $0.281 output, $0.003 cached input.
  • Hy4 (the flagship): $0.834 input, $2.501 output, $0.042 cached input.
  • OpenRouter, tencent/hy4-flash: $0.035 input, $0.1405 output, $0.0015 cached input, with a 50% discount flag on the Tencent endpoint. That is exactly half the list price; Tencent gives a November 30 end date for its own TokenHub discount, and OpenRouter does not show an end date.

For comparison, OpenRouter lists Claude Haiku 5.5 at $0.10 input and $0.50 output for prompts under 100,000 tokens, rising to $0.50 input and $2.50 output for prompts of 100,000 tokens or more. DeepSeek V4.1 Flash is listed at $0.051 input and $0.567 output. At list price and short prompts, Hy4 flash is 30% cheaper than Haiku 5.5 on input and 44% cheaper on output; on long prompts the gap is much wider.

A worked example: 20 requests of 50,000 uncached input tokens and 10,000 output tokens each (1 million in, 200,000 out in total) cost about $0.126 on Hy4 flash at list price ($0.063 at the launch discount), $0.20 on Claude Haiku 5.5 and $0.164 on DeepSeek V4.1 Flash, using the OpenRouter prices above. Send the same tokens as ten 100,000-token prompts and Haiku 5.5 moves to its long-prompt rate, about $1.00, while the other two stay the same. Reasoning is on by default for Hy4 flash on OpenRouter, and reasoning tokens are billed as output, so a model that thinks longer can erase a per-token price lead; Tencent says reinforcement learning cut Hy4 flash's average end-to-end output by 50%.

How good is Hy4 flash at coding?

Tencent publishes a table with cost per task in yuan for Hy4 flash, Hy4, Hy4 preview, DeepSeek-V4.1-Flash-1017, Qwen3.8-Flash-Next, GLM-5.3-Flash and GPT-6 Luna, each at its highest available reasoning setting. Claude is not in that table. The coding and agent rows, Hy4 flash first:

  • Terminal-Bench 4.0: Hy4 flash 19.2 at ¥0.558 per task. DeepSeek V4.1 Flash 40.0 (¥3.882), GLM-5.3-Flash 35.4, Qwen3.8-Flash-Next 29.5, Hy4 44.4, GPT-6 Luna 13.3.
  • DeepSWE v1.1 (run with mini-SWE-agent in an 8-core, 16 GB sandbox): Hy4 flash 64.6 at ¥0.290. DeepSeek V4.1 Flash 73.8 (¥1.304), GLM-5.3-Flash 67.8, GPT-6 Luna 67.1, Qwen3.8-Flash-Next 65.1, Hy4 74.9.
  • ProgramBench (almost solved): Hy4 flash 16.0. DeepSeek V4.1 Flash 29.0, Qwen3.8-Flash-Next 22.0, GLM-5.3-Flash 13.0, GPT-6 Luna 6.0, Hy4 30.0.
  • OSWorld 2.1 (computer use, partial score): Hy4 flash 68.1, DeepSeek V4.1 Flash 68.4, Hy4 78.9.
  • AutomationBench v1.0.6: Hy4 flash 52.2 at ¥0.038, DeepSeek V4.1 Flash 54.3 at ¥0.620.

The pattern: on computer use, automation and tool-assisted search Hy4 flash is close to DeepSeek V4.1 Flash at a fraction of the cost per task. On Terminal-Bench it scores less than half as much (19.2 against 40.0), and on ProgramBench about half. Tencent's own guidance matches this: it recommends Hy4 flash for deep search, document processing, slides, simple frontend demos and computer use, and Hy4 for complex, long-horizon software engineering.

Is Hy4 flash as good as Claude Haiku 5.5?

The only Claude comparison is a blind test Tencent ran with 85 of its internal WorkBuddy users on 259 office tasks in 27 categories, with external models on their high tier. Average satisfaction out of 4: DeepSeek V4.1 Flash 3.09, Claude Haiku 5.5 2.81, Hy4 flash 2.79, GLM-5.3-Flash 2.53. Median end-to-end time was 6.7 minutes for Hy4 flash and 6.8 for Haiku 5.5, and cost per task was ¥0.26 against ¥4.73.

So the launch line, satisfaction close to Haiku 5.5, holds, but the same chart has DeepSeek V4.1 Flash rated highest, at ¥1.20 per task and 10.3 minutes. These were office tasks rated by Tencent's internal WorkBuddy users, not coding tasks, and the WorkIndex benchmark is Tencent's own. No independent benchmark had been published when we checked.

How to try Hy4 flash

  • Hosted: OpenRouter model ID tencent/hy4-flash. The only provider is Tencent (FP8). It supports tools, structured outputs and reasoning; tool_choice works with none and auto but not required. Output is capped at 65,536 tokens (Haiku 5.5: 128,000).
  • Tencent's own API: Tencent Cloud TokenHub, with the 50% launch discount through November 30.
  • Self-hosting: Tencent's vLLM recipe uses the image vllm/vllm-openai:hy4-flash with the FP8 weights and --tensor-parallel-size 8; the SGLang image lmsysorg/sglang:hy4-flash uses --tp 8. That is an eight-GPU server, not a laptop. Transformers support is not merged yet; the card says to install from a pull-request branch.
  • A GGUF build for end-to-end testing is published under AngelSlim/Hy4-flash-GGUF.

Try it with your AI. Want to see whether Hy4 flash can take over cheap, repetitive steps in your own pipeline? Paste this into your coding agent. It runs a small side-by-side test against the model you use now and touches nothing else.

I want to test whether Tencent's Hy4 flash (OpenRouter model ID tencent/hy4-flash) can
replace [MY CURRENT MODEL ID] for one low-risk, repetitive step in my project: [DESCRIBE
THE STEP, e.g. summarising failing test logs].

First check: that an OpenRouter API key is already available as the environment variable
OPENROUTER_API_KEY (do not print it, do not ask me to paste it, do not write it to any
file). If it is missing, stop and tell me.

Then:
1. Pick [NUMBER, e.g. 10] real inputs for this step from [PATH OR SOURCE]. Show me the
   list and wait for my OK before sending anything, because the inputs will leave my
   machine.
2. Write a small script that sends each input to both models through
   https://openrouter.ai/api/v1/chat/completions with the same prompt. For Hy4 flash set
   temperature 0.9 as Tencent recommends (OpenRouter does not list top_p for this
   model). Reasoning is on by default; keep the reasoning setting the same for every run
   and tell me which one you used. Record output, prompt tokens, completion tokens
   (reasoning included) and the usage cost OpenRouter returns. Keep each prompt under
   100,000 tokens unless I say otherwise, because Claude Haiku 5.5 is billed at a higher
   rate above that.
3. Save results to a new folder [OUTPUT FOLDER]; do not modify or delete existing files.
4. Give me a table: input, both outputs side by side, tokens, cost. Do not judge quality
   for me beyond flagging outputs that are empty, truncated or ignore the instructions.

Do not install new packages without asking. Do not use this for multi-step coding tasks;
Tencent recommends its larger Hy4 model for long-horizon software engineering.

Known limits

  • Every benchmark above comes from Tencent. Treat them as the vendor's numbers until independent results appear.
  • Weak on long terminal coding by Tencent's own table (Terminal-Bench 4.0: 19.2).
  • The OpenRouter price is currently half the list price with no end date shown; budget with the list price.
  • Licence: the extra terms apply to model-as-a-service businesses, meaning you give third parties remote access to inference or fine-tuning, for example through an API. End-user products that confine the model to specific features or workflows, and services that only relay requests to a third-party model service, are excluded. A MaaS business needs a separate written licence from Tencent if its worldwide revenue exceeds US$50 million in any 12-month period, or before using the model in a product with more than 100 million monthly active users. Purely internal use, where neither the model nor its outputs reach third parties, is exempt. Licensing contact: [email protected].
  • The total parameter count differs between Tencent's pages (220B on Hugging Face, 224B on the announcement).

FAQ

Is Hy4 flash open source? Its weights are open under the Tencent Hy License 1.0, which is based on Apache 2.0 but adds commercial terms for large model-as-a-service businesses, so it is not the standard Apache licence.

Is Hy4 flash free? It is free inside Tencent's WorkBuddy app through November 30. The API is paid: $0.070 input and $0.281 output per million tokens at list price, half that during the launch discount.

Can I run Hy4 flash locally? The official vLLM and SGLang recipes use eight GPUs with FP8 weights. A GGUF version exists for testing, but Tencent has not published hardware requirements for it.

Does Hy4 flash support tool calling? Yes. OpenRouter lists tools and structured outputs (tool_choice none or auto, not required), and Tencent's vLLM recipe enables a hy_v4 tool-call parser.

Source: Hy4 flash announcement (Tencent Hy), tencent/Hy4-flash model card (Hugging Face), Tencent Hy License 1.0 (Hugging Face), Tencent-Hunyuan/Hy4-flash (GitHub), tencent/hy4-flash (OpenRouter), Claude Haiku 5.5 (OpenRouter), DeepSeek V4.1 Flash (OpenRouter)

This story first appeared in the claudefox daily briefing for November 18, 2026.