Vals put Hy4 Preview first among open-weight models on code migration at $3.41 a test

Cost per test is a measurement of a harness as much as of a model. Vals reached Hy4 Preview through OpenRouter with temperature at 0.9, Top P at 1 and a maximum output length of 64,000 tokens ^[8]. Tencent lists API access at $0.834 per million input tokens and $2.501 per million output tokens ^[12]. Under that output cap, one full-length response bills at most about $0.16 of output ^[21]. If the whole $3.41 Code Migration bill were output at the cap, it would take at least 21 maximum-length responses to spend it ^[24].

Which side of the meter the money went to decides whether the price transfers. At Tencent's list rates, $3.41 also buys about 4.09 million input tokens ^[19]. An agent loop that resends a growing transcript every turn gets to that total fast, and the turn count is a property of the scaffold. The Code Migration test also cost about 2.7 times Hy4 Preview's average test across the whole index ^[20].

Temperature 0.9 with Top P at 1 ^[8] is not the sampling setting I would choose for a comparison suite on code. Repeated runs diverge at that temperature, and a run that diverges into a failure costs another attempt. RuntimeWire notes that cost per test stays tied to the evaluator's harness and task mix, and that different providers, prompts and agent scaffolds produce different economics ^[9].

For $3.41 to be your number, the job has to be Vals' job, reimplementing a working program in another language ^[3], your loop has to make roughly the same number of turns, and your bill has to come from whichever OpenRouter provider Vals bought from ^[8].

Tencent presented Hy4 Preview on August 28th as a productivity flagship for coding, office work and scientific research ^[13]. The Hy team says the training data was developed alongside Tencent specialists in software engineering, game development, finance and security, and that the model was co-designed with products including CodeBuddy and WorkBuddy ^[14]. The Vals spread follows that provenance. It is strongest where the task is fully specified ^[3], and between the extremes sit 10th of 61 on Legal Research Bench and 13th of 62 on Harvey's Legal Agent Benchmark ^[15].

The weights are Apache 2.0, with deployment instructions covering vLLM and SGLang ^[11]. Self-hosting changes what you are buying. 49 billion of the 770 billion parameters activate per token, or 6.4% ^[18], so compute per token tracks the 49 billion while the weights you hold resident track all 770 billion.