Current model · Tencent
Hy4 benchmarks, pricing and specs
Hy4 is a Tencent model released on November 4, 2026 that takes text and image input and produces text. It costs $0.834 per 1M input tokens and $2.501 per 1M output tokens.
- Per 1M input tokens
- $0.834Listed today
- Per 1M output tokens
- $2.501Listed today
Data verified November 5, 2026 · Dataset 2026-11-05
Specs
Key facts
- Provider
- Tencent
- Model family
- Hy4
- Release date
- November 4, 2026
- Status
- Current
- Access
- API only
- API model id
- hy4
- Context window
- Not reported
- Max output
- Not reported
- Input
- text, image
- Output
- text
- Task families
- text, multimodal
- Declared capabilities
- reasoning, agentic coding, image understanding, chart understanding
Charts
Against every current text model
Where Hy4 lands among the current models of its task, on every independent benchmark, on price and on context.
Where it stands
20 of 55 list a lower price per 1M input tokens. Each bar is the share of current models it beats.
Price and context
Price per 1M input tokens
#21of 55
$0.834 · cheaper than 63%
Price per 1M output tokens
#20of 55
$2.501 · cheaper than 65%
Benchmarks
Independent benchmark results
Scores come from independent evaluators that ran the model themselves, pinned to a dated snapshot. Rank is among current models in the same snapshot. Tencent's own launch figures are in the next section, because they are not measured the same way.
No independent benchmark has published a result for Hy4 yet. It is still compared below on price, context and capabilities.
Provider results
Reported by Tencent
Tencent's own results for Hy4, beside the rivals it chose, as its table lists them. Labs pick the benchmarks and run them their own way, so compare within one table only. Internal benchmarks and other evaluators' indexes are left out.
Hy4 launch table, November 4, 2026Tencent · 2026-11-05
| Benchmark | Hy4 | GLM-5.3 | DeepSeek-V4.1-Flash | Kimi K3.1 Pro | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | 74.9% | 68.1% | 73.8% | 76.3% | 72.7% | 72.6% |
| SWE-Marathon v1.1 | 43.8% | 35.6% | 46.3% | 53.8% | 39.4% | 44.4% |
| ProgramBenchalmost solved | 30.0% | 18.0% | 29.0% | 24.5% | 46.0% | 56.5% |
| Terminal-Bench 4.0 | 44.4% | 42.9% | 40.0% | 66.8% | 55.5% | 66.4% |
| FrontierCode 2.0 | 31.0% | 21.6% | 28.5% | 25.5% | 40.6% | 46.4% |
| APEX-Agents v1.1pass@1 | 75.2% | 64.3% | 72.9% | 68.2% | 61.7% | 74.8% |
| Agents' Last Exam v1.1pass rate | 41.0% | 26.2% | 35.3% | 43.4% | 40.2% | 43.4% |
| AutomationBench v1.0.6 | 56.2% | 54.1% | 54.3% | 65.6% | 51.4% | 58.8% |
| HLE-Fullwith tools | 59.1% | Not listed | 57.1% | 61.8% | 60.6% | 66.2% |
| Terminal-Bench-Science 0.1 | 19.8% | 10.0% | 13.7% | 41.7% | 60.9% | 56.2% |
| PostTrainBench v1.2 | 35.2% | 31.6% | 29.4% | 36.2% | 38.2% | 44.2% |
| OSWorld 2.1full, partial score | 78.9% | Not listed | 68.4% | 79.6% | 75.2% | 81.8% |
| MMMU-Pro | 78.6% | Not listed | 76.7% | 83.4% | 84.9% | 88.2% |
| ZeroBench mainwith tools, pass@5 | 55.4% | Not listed | 51.0% | 56.9% | 65.3% | 63.5% |
Tencent's appendix to the Hy4 announcement, every model at its highest reasoning setting. Where Tencent lists a lab's reported score beside its own run, the score Tencent measured is shown. DeepSeek-V4.1-Flash is the 1017 build. Left out here: Tencent's internal benchmarks, the Hy4 preview column and GLM-5.3 on HLE, which Tencent ran on the text-only subset.
Fit
Strengths, cautions and best uses
- Strengths
- Tencent's flagship and the successor to the open Hy4 preview (a 770B-parameter mixture-of-experts, 49B active per token): it adds image understanding and, for the first time, computer use across command-line and graphical apps
- In Tencent's runs it scores 74.9% on DeepSWE v1.1 and 75.2% on APEX-Agents v1.1, against 72.6% and 74.8% for Claude Opus 5.5
- $0.834 input, $2.501 output and $0.042 cached input per 1M tokens, the same as Hy4 preview
- Cautions
- API only for now: Tencent says it will publish the weights and a system card later, and only Hy4 preview (August 2026, Apache 2.0) is open today
- In Tencent's own table it scores 44.4% on Terminal-Bench 4.0 against 66.8% for Kimi K3.1 Pro, and Tencent says a clear gap remains to top proprietary models on research-level math, long-horizon coding and hard visual questions
- Tencent publishes no context window or output limit for Hy4 (the preview card lists 1M tokens), and no independent benchmark score is recorded here yet
- Best for
- Agents that work across terminals, desktop apps and browsers on a tight budget
- Finance and office work: research reports, spreadsheets and slides
From benchmark to production
Building on Hy4?
Public scores rarely hold on your own data. We test it against the alternatives on your real tasks, then build the agent or pipeline around the model that wins.
AI Model Cost AuditLaunch price $100, regular price $150What the audit includes
Alternatives
Alternatives to Hy4
The closest current text models by price. The cheapest model that scores as well or better is listed first; failing that, the cheapest one within 3% of the benchmark's scale.
On score and price
Hy4 beside the 4 alternatives listed below, on Arena Text and on blended price per 1M tokens.
Arena Text
Measured October 28, 2026, higher is better
- Hy4
- Not scored
- Gemini 3.8 Flash
- 1,496
- Mistral Large 4
- 1,432
- GPT-6.1 Terra
- Not scored
- Step 5 Preview
- 1,448
Blended price per 1M tokens
Listed today, lower is better
- Hy4
- $1.25
- Gemini 3.8 Flash
- $1.50
- Mistral Large 4
- $1.05
- GPT-6.1 Terra
- $2
- Step 5 Preview
- $1.43
Gemini 3.8 Flash
- 10% cheaper per 1M input tokens, $0.75 vs $0.834
- From Google
Mistral Large 4
- 16% cheaper per 1M input tokens, $0.70 vs $0.834
- From Mistral AI
GPT-6.1 Terra
- 20% more per 1M input tokens, $1 vs $0.834
- From OpenAI
Step 5 Preview
- 20% more per 1M input tokens, $1 vs $0.834
- From StepFun
Further reading
More on Hy4
FAQ
Hy4 FAQ
How much does Hy4 cost?
Hy4 costs $0.834 per 1M input tokens and $2.501 per 1M output tokens at the standard tier. It is the #21 cheapest of 55 current text models priced per 1M input tokens. Source: Tencent, verified November 5, 2026.
What benchmark results has Tencent published for Hy4?
In Hy4 launch table (November 4, 2026), Tencent reports 74.9% on DeepSWE v1.1, 43.8% on SWE-Marathon v1.1, 30.0% on ProgramBench (almost solved) and 44.4% on Terminal-Bench 4.0. These are Tencent's own figures, beside the rivals it chose, not independent measurements.
Is Hy4 open source or API only?
Hy4 is available through Tencent's API only; its weights are not published.
What are the best alternatives to Hy4?
The closest current alternatives by price and measured performance are Gemini 3.8 Flash (10% cheaper per 1M input tokens, $0.75 vs $0.834), Mistral Large 4 (16% cheaper per 1M input tokens, $0.70 vs $0.834), GPT-6.1 Terra (20% more per 1M input tokens, $1 vs $0.834) and Step 5 Preview (20% more per 1M input tokens, $1 vs $0.834).
Sources
Sources
- Introducing Hy4 · Tencent · official · verified November 5, 2026
- Hy4 preview model card · Tencent · official · verified November 5, 2026
All 241 sources behind the benchmarks, with their versions, licenses and notes, are in the source catalog.
Cite this page
Van Data Team. "Hy4 Benchmarks, Pricing and Specs." AI Model Benchmark Dataset 2026-11-05. https://vandatateam.com/ai-model-benchmarks/hy4
Reuse the data: JSON · CSV · or explore every model in the AI model benchmark comparison.
Production evaluation
Deciding whether Hy4 fits your workload?
Public benchmarks narrow the shortlist. We test the finalists on your own tasks, data and cost limits, then build the routing and review controls to run the winner in production.
- Evaluation cases and acceptance thresholds tied to real work
- Model routing, fallbacks and cost controls