Hy4 Preview vs Qwen3.8-Max vs GLM-5.3: Open Frontier
The Short Answer
Three open-weight frontier models arrived within four weeks of each other in August 2026. Tencent Hy4 preview (August 28) is the cheapest at $0.834/$2.501 per MTok. GLM-5.3 (August 14, weights published late August) is the strongest declared coder at $1.40/$4.40. Qwen3.8-Max (August 3) is the largest at 2.4 trillion parameters and $2/$6.
If you want one rule: Hy4 for cheap volume, GLM-5.3 for hard coding, Qwen3.8-Max for long-horizon autonomous work.
Head-to-Head
| Hy4 preview | GLM-5.3 | Qwen3.8-Max | |
|---|---|---|---|
| Vendor | Tencent (Hunyuan) | Z.ai (Zhipu) | Alibaba (Qwen) |
| Released | Aug 28, 2026 | Aug 14, 2026 | Aug 3, 2026 |
| Parameters | 770B total / 49B active | ~743B base, post-trained | 2.4T |
| Context | >1,000,000 | 1,000,000 | 1,000,000 |
| Input / MTok | ~$0.834 | $1.40 | $2.00 |
| Output / MTok | ~$2.501 | $4.40 | $6.00 |
| Cached input | — | $0.26 | $0.25 |
| 30K/5K task | $0.038 | $0.064 | $0.090 |
| Weights | Hugging Face | Hugging Face, custom licence | 2.4T custom; 27B Apache 2.0 |
| Best at | Cheap volume, orchestration | Agentic coding, cyber defence | Long-horizon autonomy |
Last verified: August 30, 2026.
Price: Hy4 Wins, But Check Who You Are Comparing To
On the reference 30,000-input / 5,000-output task, Hy4 preview costs $0.038. GLM-5.3 is 68% more expensive at $0.064. Qwen3.8-Max is 137% more expensive at $0.090.
Against closed frontier models the gap is much larger. Claude Opus 5 at $5/$25 costs $0.275 per task; GPT-5.6 Sol at $4/$20 costs $0.22. Hy4 preview is roughly one-seventh of Sol.
One important caveat: the cheapest option in this space is not on the list above. DeepSeek V4 Pro at $0.66/$1.98 off-peak runs $0.0297 per task — cheaper than Hy4. But DeepSeek repriced to peak/off-peak on August 16, 2026, and peak (01:00–04:00 and 06:00–10:00 UTC) doubles that to $0.0594. If your load sits in Asia-Pacific business hours, Hy4’s flat rate may well beat it.
The Licence Story Is the Real 2026 Change
“Open weights” stopped meaning one thing this year, and August made that unmistakable.
GLM-5.3 published its weights to Hugging Face — but dropped the MIT licence Z.ai had used previously. The new terms require providers above $10 billion in revenue to undergo a security review. That is a licence aimed squarely at hyperscalers, not at you. For a startup or an enterprise below that threshold it changes nothing practical; for AWS, Azure or Google Cloud it is a gate.
Qwen3.8-Max split its release: the 2.4T checkpoint ships under a custom licence, while Qwen3.8-27B is genuine Apache 2.0. The 2.4T open-weight release is also text-only — no vision, no native 1M context — which is a meaningful downgrade from the hosted endpoint.
Hy4 preview published weights on Hugging Face as part of the launch itself, with open-sourcing framed as the headline rather than a follow-up.
The lesson: read the licence file, not the press release. “Open-weight” in late 2026 covers everything from MIT to revenue-gated source-available.
Capability: What Each One Actually Claims
GLM-5.3 makes the sharpest claim: Z.ai says it is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench, and markets it for agentic coding and cyber defence. In-house benchmark, so discount accordingly — but the positioning is unambiguous.
Qwen3.8-Max is built for autonomous, multi-day coding and research work at 2.4 trillion parameters. Alibaba’s own framing at launch was that it ranks “second only to Fable 5” — a bold claim it notably did not back with a published benchmark table at release.
Hy4 preview takes a different angle. Tencent emphasises agentic research orchestration: Hy4 coordinating several Codex sessions in parallel and evaluating their results. It is positioned as a supervisor of other agents, and its training data was built with internal Tencent domain experts around work they actually ship.
Notice that all three headline claims are vendor-internal. As of August 30, 2026 none of these models has a settled independent benchmark picture. That is the honest state of the field.
Which Should You Pick?
Choose Hy4 preview if cost per token dominates your economics, you are running high-volume agentic pipelines, or you want a cheap orchestrator coordinating more expensive specialist models. Accept that it is a preview and the benchmark evidence is thin.
Choose GLM-5.3 if your workload is hard multi-file coding or security work, and you want the model with the most specific coding claim plus published weights. Verify the licence terms apply cleanly to your company size.
Choose Qwen3.8-Max if you are running long-horizon autonomous tasks — multi-day research or engineering runs — where raw capacity matters more than per-token price, and you want a hosted endpoint with vision and full 1M context.
Choose none of them if you need verified frontier accuracy on high-stakes output today. Claude Opus 5 and GPT-5.6 Sol cost 3–7x more per task and, as of August 2026, still carry the stronger independent evidence base.
The Pattern Worth Copying
Do not pick one. Route.
Send bulk classification, extraction and summarisation to the cheapest tier that passes your evals. Escalate the hard 5–10% to a frontier model. Teams running this pattern typically land blended costs 60–80% below single-model frontier deployment, because most tokens in most production systems are genuinely easy.
The August 2026 releases make that easier, not harder: there are now three credible open-weight options in the sub-$0.10-per-task band, all with published weights and 1M context windows.