Tencent launches Hy4 Flash, a high-speed, low-cost model that matches Claude Haiku 5.5 performance while delivering an 18x reduction in task costs.

On November 16, 2026, Tencent Hunyuan officially announced the release of Hy4 Flash, a specialized language model designed to redefine the economics of everyday AI productivity. As the industry shifts from pure reasoning benchmarks to real-world deployment efficiency, Hy4 Flash arrives as a strategic answer to the rising costs of high-frequency LLM tasks.
Positioned as the most cost-effective model in its quality tier, Hy4 Flash is engineered for developers and enterprises that require reliable performance for high-volume workflows without the premium price tag of frontier flagship models. It bridges the gap between massive, high-latency models and lightweight, low-intelligence models, offering a 'sweet spot' for agentic workflows and professional tools.
Hy4 Flash is part of the broader Hy4 model family, which includes the massive 770B parameter open-weight Mixture-of-Experts (MoE) base model. While the base model provides the heavy-duty reasoning capabilities, Hy4 Flash is optimized for speed and throughput, utilizing a highly efficient architecture that balances activated parameters with total parameter count.
The architecture is designed to support multimodal reasoning and computer use, allowing the model to interpret complex visual data and interact with digital interfaces. This makes it more than just a text generator; it is a functional component for autonomous agents capable of executing multi-step processes in professional software environments.
In human blind tests, Hy4 Flash has demonstrated performance that matches Claude Haiku 5.5 (high setting) on standard daily work tasks. This parity in qualitative output is particularly impressive given the drastic difference in operational costs. For developers building RAG pipelines or automated customer support, this represents a massive leap in ROI.
The flagship Hy4 base model also sets a high bar for the series, ranking #1 on the WorkBuddy benchmark with an average score of 83.18 and a remarkable 96.15% pass rate. This places it in direct competition with top-tier models like GLM-5.3 and Kimi K3. The evaluation methodology for these benchmarks focuses on practical utility, including reasoning accuracy, instruction following, and long-context retrieval integrity.
Tencent has implemented an aggressive pricing strategy to capture the high-frequency task market. Hy4 Flash is significantly cheaper than its counterparts; for instance, the base Hy4 on OpenRouter costs $0.42/$1.25 per million tokens, whereas Flash is roughly 6x cheaper on input and 4.5x cheaper on output.
A standout feature for developers is the highly optimized cached input pricing. By utilizing context caching, developers can reduce their input costs to a fraction of the standard rate, making long-context RAG and repetitive agentic prompts economically viable for the first time at this scale.
Hy4 Flash is specifically targeted at three primary domains: everyday productivity, professional tool integration, and long-horizon task execution. Its low latency and low cost make it the ideal engine for autonomous agents that need to 'think' through dozens of steps to complete a single user request.
Common implementation patterns include high-volume RAG (Retrieval-Augmented Generation) systems, automated coding assistants, and real-time data extraction from unstructured documents. Its ability to handle computer use also opens doors for RPA (Robotic Process Automation) 2.0, where the LLM acts as the primary interface for legacy software.
Developers can access Hy4 Flash immediately via the Tencent Cloud API. The model is served through the official endpoint at hy.tencent.ai, where users can manage API keys, monitor usage, and integrate the model into their existing tech stacks using standard RESTful requests.
For those interested in the larger-scale reasoning capabilities, the Hy4 base model weights are available on Hugging Face, allowing for local deployment and fine-tuning for specialized industrial applications.
API Pricing β Input: $0.070 per million tokens / Output: $0.281 per million tokens / Context: $0.003 per 1M tokens (cached input)