Tencent HY4 Preview vs tencent-hy3: Six Times the Price, and What the Leap Actually Buys

Tencent HY4 Preview is the first model in the Hunyuan family that makes its own predecessor look cheap on purpose: Tencent lists it at six times tencent-hy3's input price and four and a half times its output price, and the launch deck argues the premium is earned. Released and open-sourced on August 28, 2026, Tencent HY4 Preview reports a software-engineering score on DeepSWE that more than doubles Hy3's 28.0, a terminal-driving 85.4 on Terminal-Bench 2.1, and a context window that jumps from 256K to over a million tokens. tencent-hy3, which went official on July 6, is the mature workhorse of the family — 295B total parameters, 21B active, a quarter of the context, and a price tag closer to a rounding error. This is not a contest the spec sheet leaves close. The question is whether the gap is worth the money for what you actually run, and the answer splits by workload.

The scoreboard: where the two actually diverge

Every benchmark in Tencent's materials is vendor-reported — run on Tencent's own evaluation setups at each model's launch, unreproduced by an independent lab, and for the flagship the model has been public for less than a day. Treat the scores as the ceiling Tencent believes it hit, and weight the pattern more than any single digit. The pattern is unmistakable, and it is concentrated on agentic engineering work rather than general knowledge:

• DeepSWE — Tencent HY4 Preview: 64.3. tencent-hy3: 28.0. This is the biggest single jump in the pairing, a more-than-doubling on a repository-scale software-engineering benchmark, and it is the number that separates a chat model from a model you hand a whole codebase.

• Terminal-Bench 2.1 — Tencent HY4 Preview: 85.4, which Tencent says ties Claude Opus 5 and passes DeepSeek V4 Pro. tencent-hy3: 71.7 as reported at its July launch. Driving a real terminal to completion on multi-step tasks is exactly the kind of long-horizon work Hy3 was weakest at.

• Toolathlon-Verified — Tencent HY4 Preview: 74.1, which Tencent says surpasses Qwen 3.8 Max and GPT-5.6 Sol. No comparable Hy3 figure was published.

• Internal blind test — 163 experts scored 203 engineering tasks at 2.99/4.00 for Tencent HY4 Preview, edging GLM-5.3's 2.92 and Kimi K3's 2.94. This is Tencent's reviewers on Tencent's tasks; treat it as directional.

The consistent thread: the gains are on terminal driving, tool calling, and repository-scale tasks — the productivity positioning Tencent has been pushing since Hy3, now with the compute to back it up.

The price premium, line by line

Here is where the comparison stops being about bragging rights. Tencent's list prices, per million tokens:

• Input — Tencent HY4 Preview: 6 yuan (~US$0.85). tencent-hy3: 1 yuan (~US$0.14). Six times.

• Output — Tencent HY4 Preview: 18 yuan (~US$2.50). tencent-hy3: 4 yuan (~US$0.56). Four and a half times.

• Cache hit — Tencent HY4 Preview: 0.3 yuan. tencent-hy3: 0.25 yuan. Nearly the same, which matters more than it looks, because agentic loops re-send the same system prompt and conversation prefix constantly.

Run a realistic agentic-coding workload through the blended number — say 20 million input and 4 million output tokens in a month, a busy single-developer agent budget. Tencent HY4 Preview: 120 yuan plus 72 yuan, about 192 yuan (~US$27). tencent-hy3: 20 yuan plus 16 yuan, about 36 yuan (~US$5). The flagship costs five-plus times as much to run at the same token mix — and it will ask for more output tokens per task, because it thinks longer.

That is not a reason to avoid Tencent HY4 Preview; the DeepSWE jump is exactly the kind of capability a premium exists to buy. It is a reason to price the workload before you route to it. And it is the place the routing layer earns its keep: tencent-hy3 is already on OrcaRouter at the provider's list price — 0% markup, so the number you see is the serving provider's own price, not a resold and marked-up version of it — with automatic failover across providers, and a vendor price change goes live on our side the same day.

Four times the context, double the active compute

• Architecture — Tencent HY4 Preview: 770B total, 49B active MoE. tencent-hy3: 295B total, 21B active. The flagship puts roughly 2.3 times as much compute on every token.

• Context window — Tencent HY4 Preview: over 1M tokens. tencent-hy3: 256K. A full repository or a batch of financial documents fits in the flagship's window as a single request; on Hy3 the same job needs chunking, retrieval, or an agent loop.

• Reasoning control — tencent-hy3 ships a per-request reasoning_effort setting (no_think, low, high) plus a speculative-decoding layer that keeps it fast. Tencent HY4 Preview, by Tencent's own admission, tends toward long thinking before first output and over-self-verifies on complex tasks — burning tokens double-checking work it has already done.

That last line is the sleeper difference for latency-sensitive use. Hy3 can be told to answer directly; Tencent HY4 Preview, at least in this early form, often cannot help but think. For a low-latency chat or a high-throughput extraction pipeline, the flagship's extra capability is wasted and its extra thinking is a tax. For an offline batch engineering job, the tax is exactly what buys the better answer.

The preview tax: what Hy3 still has

"Preview" is in the name of Tencent HY4 Preview and it behaves like it. There is no vision or multimodal input — the model is text-only, so anything image- or video-adjacent stays on whatever model you already use for that. The two-week free trial on WorkBuddy and CodeBuddy is a taster, not a plan. Tencent has been explicit that this is an early iteration of Hy4, with pre-training and post-training improvements still to come on a cadence that has produced a major version roughly every two months since February. Independent benchmarks for the flagship: none yet, anywhere.

tencent-hy3 is the opposite of all of that. It is an official, stable release that has been in production since July. Its free tier runs to September 30. Its reasoning levels give you a dial rather than a temperament, and its speculative-decoding layer and 21B active parameters make it the cheap, fast, predictable half of this family — the model you point at the default path and forget about.

Who should pick which

Pick Tencent HY4 Preview when the job is the point — a hard software-engineering task, a repository-scale context, an agentic workflow where a better answer justifies five times the token bill and you can afford the latency of a model that thinks before it speaks. Pick tencent-hy3 when the job is the constraint — high throughput, low latency, a tight budget, or any task where you want the model to answer rather than deliberate.

The practical pattern for a pairing like this is escalation: route the cheap, controllable model on the default path and escalate to the flagship only when the task demands it. That is what OrcaRouter's routing DSL is for — composing several models into one call so the policy is configuration rather than code — and automatic failover means Hy3's path keeps serving even when one provider degrades. OrcaRouter does not route Tencent HY4 Preview yet; Tencent has not opened the model to third-party inference, so for now it runs on the vendor's own Tencent Cloud TokenHub endpoint and on self-hosted deployments of the open weights. tencent-hy3, meanwhile, is in the catalog today at the provider's list price — and the day the flagship opens up, it slots into the same routing layer the same way, turning the swap from one model to the other into a one-line change instead of a rewrite.

The verdict is boring in the best way: the flagship is worth the premium for exactly the work it is priced to do, and the workhorse is still the right call for everything that just needs to get done. The six-times price gap is real, the 28-to-64 DeepSWE gap is real, and neither cancels the other — you just have to know which one you are buying.