DeepSeek V4 Pro 0423 vs Hy4 preview
Hy4 preview leads DeepSeek V4 Pro 0423 on coding, 44.3 to 38.7 on the Vibe Coding Index, at about the same output price ($2.501 vs $2.976 per million). Updated October 22, 2026; 17 benchmarks have a result for both.
DeepSeek V4 Pro 0423 or Hy4 preview: which to pick
Pick Hy4 preview for coding: it scores 44.3 on the Vibe Coding Index against 38.7 for DeepSeek V4 Pro 0423, though on 3 benchmarks against 5.
Pick DeepSeek V4 Pro 0423 for work measured by Terminal-Bench 4 (Vals AI), where it leads.
- Shared results
- 17
- DeepSeek V4 Pro 0423 only
- 67
- Hy4 preview only
- 7
- Like-for-like categories
- 5 of 9
Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, six public coding benchmarks
Hy4 preview
Hy4 preview scores 44.3 against 38.7.
5 and 3 benchmarks behind the scores
Intelligence
AA Intelligence Index
No call
Only DeepSeek V4 Pro 0423 has a AA Intelligence Index result, 30.4.
Directional only
Agentic work
AA Agentic Index
No call
Only DeepSeek V4 Pro 0423 has a AA Agentic Index result, 26.3.
Directional only
Reasoning
GPQA Diamond
No call
Only DeepSeek V4 Pro 0423 has a GPQA Diamond result, 89.4%.
Directional only
Knowledge
MMLU Pro
No call
Only DeepSeek V4 Pro 0423 has a MMLU Pro result, 87.3%.
Directional only
Web design
Design Arena: websites
No call
Only DeepSeek V4 Pro 0423 has a Design Arena: websites result, 1234.
Directional only
General
Vals Index
No call
Only Hy4 preview has a Vals Index result, 53.2%.
Directional only
Long documents
Listed context window
No call
Both take 1.05M tokens.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Hy4 preview
DeepSeek V4 Pro 0423 $0.0028, Hy4 preview $0.0021 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only DeepSeek V4 Pro 0423 has the number, $0.57.
Directional only
At a glance
- Coding index
- DeepSeek V4 Pro 0423 38.7 from 5 benchmarks, Hy4 preview 44.3 from 3 benchmarks.
- Widest category
- Math, 59.0 points apart on average over 1 shared benchmark.
- Widest single gap
- ProofBench: DeepSeek V4 Pro 0423 16.0% against Hy4 preview 75.0%, 59.0 points apart on a 0 to 100 scale.
- Output price
- DeepSeek V4 Pro 0423 $2.976 against Hy4 preview $2.501 per million output tokens, 1.2x apart.
- Context window
- Both take 1.05M tokens.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | DeepSeek V4 Pro 0423 | Bars | Hy4 preview | Ties |
|---|---|---|---|---|
| Coding | 1 | 6 | 1 tie | |
| Agentic work | 0 | 4 | ||
| Math | 0 | 1 | ||
| Knowledge | 0 | 2 | ||
| General | 0 | 2 | ||
| All shared boards | 1 | 15 | 1 tie |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index16
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| LMArena WebDev | 1447 | 1631 | Hy4 preview |
| AA Coding Index | 59.4 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LiveCodeBench (Vals AI) | 87.5% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| SWE-bench Verified (Vals AI) | 77.4% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Terminal-Bench 4 (Vals AI) | 11.1% | 8.1% | DeepSeek V4 Pro 0423 |
| Vibe Code Bench (Vals AI) | 49.9% | 77.5% | Hy4 preview |
| Code migration (Vals AI) | 26.2% | 47.4% | Hy4 preview |
| ProgramBench (Vals AI) | 0.0% | 0.0% | Tie |
| IOI (Vals AI) | Not measured | 59.3% | Only Hy4 preview measured |
| Terminal-Bench 2.1 (Vals AI) | 50.2% | 55.1% | Hy4 preview |
| SkillsBench (Vals AI) | 51.3% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| CyberBench (Vals AI) | Not measured | 65.4% | Only Hy4 preview measured |
| FrontierCode | 17.6% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| WeirdML | 48.9% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| ALE-Bench | 1006.1 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: coding | 1470 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: software and IT | 1469 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena WebDev: React | 1441 | 1653 | Hy4 preview |
| LMArena WebDev: HTML | 1435 | 1620 | Hy4 preview |
| Neither model has a result on the other 20 boards here | |||
IntelligenceDirectional onlyPublic boards, shown as published, kept out of the index2 boards, none shared
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| AA Intelligence Index | 30.4 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Epoch Capabilities Index | 149.1 | Not measured | Only DeepSeek V4 Pro 0423 measured |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index04
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| AA Agentic Index | 26.3 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: agents building web apps | 1000 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: agents building full-stack apps | 948 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: agents building Godot games | 1059 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| SREBench (Vals AI) | Not measured | 2.3% | Only Hy4 preview measured |
| Finance Agent (Vals AI) | 44.1% | 55.1% | Hy4 preview |
| Tax Agent Bench (Vals AI) | 25.2% | 32.6% | Hy4 preview |
| Legal Research Bench (Vals AI) | 23.1% | 45.2% | Hy4 preview |
| Harvey's Legal Agent Benchmark (Vals AI) | Not measured | 8.8% | Only Hy4 preview measured |
| Excel Modeling Benchmark (Vals AI) | 51.6% | 58.1% | Hy4 preview |
| BioMysteryBench (Vals AI) | Not measured | 69.3% | Only Hy4 preview measured |
| Terminal-Bench-Science (Vals AI) | Not measured | 1.4% | Only Hy4 preview measured |
| Vending-Bench 2 | $3,285 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Neither model has a result on the other 28 boards here | |||
ReasoningDirectional onlyPublic boards, shown as published, kept out of the index8 boards, none shared
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| GPQA Diamond | 89.4% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| GPQA Diamond (Epoch AI run) | 90.9% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Chess puzzles | 19.0% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Mystery game puzzles | 17.0% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: hard prompts | 1461 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: expert | 1463 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: hard prompts in English | 1466 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Kagi LLM Benchmark | 53.5% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Neither model has a result on the other 14 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| OTIS Mock AIME | 96.7% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| FrontierMath, tiers 1 to 3 | 45.3% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| FrontierMath, tier 4 | 2.4% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| ProofBench | 16.0% | 75.0% | Hy4 preview |
| LMArena Text: math | 1443 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: mathematical work | 1455 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Neither model has a result on the other 7 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| MMLU Pro | 87.3% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LegalBench (Vals AI) | 80.3% | 83.8% | Hy4 preview |
| CaseLaw (Vals AI) | 59.4% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| CorpFin (Vals AI) | 61.4% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| TaxEval (Vals AI) | 72.1% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| MedCode (Vals AI) | 40.5% | 43.3% | Hy4 preview |
| SimpleQA Verified | 47.0% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: science | 1470 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: medicine and healthcare | 1466 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: legal and government | 1464 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: business and finance | 1443 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Neither model has a result on the other 7 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index8 boards, none shared
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| Design Arena: websites | 1234 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: UI components | 1219 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: game dev | 1228 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: data viz | 1203 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: 3D | 1253 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: all code categories | 1240 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: SVG | 1139 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Design Arena: ASCII art | 1149 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Neither model has a result on the other 1 board here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| Vals Index | Not measured | 53.2% | Only Hy4 preview measured |
| Public Benefits Bench (Vals AI) | 62.9% | 68.6% | Hy4 preview |
| MedScribe (Vals AI) | 75.1% | 83.6% | Hy4 preview |
| DTBench | 90.7% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMCA | 41.2% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Surface Evolver Bench | 40.0% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| CL-bench Life | 13.5% | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text | 1451 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: instruction following | 1445 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: creative writing | 1446 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: English | 1459 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: non-English | 1439 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: Chinese | 1481 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: German | 1458 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: Japanese | 1446 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: Korean | 1409 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: French | 1473 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: Spanish | 1457 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: Russian | 1453 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: Polish | 1461 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: multi-turn | 1468 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: longer queries | 1458 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: writing and language | 1447 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| LMArena Text: entertainment and media | 1431 | Not measured | Only DeepSeek V4 Pro 0423 measured |
| Neither model has a result on the other 22 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | DeepSeek V4 Pro 0423 | Hy4 preview | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
Hy4 preview costs less
Repository review
50K input + 3K output tokens
Hy4 preview costs less
Agent loop
200K input + 10K output tokens
Hy4 preview costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | DeepSeek V4 Pro 0423 | Tencent Hy4 preview |
|---|---|---|
| Maker | DeepSeek | Tencent |
| Weights | Open weights | Open weights |
| Released | April 24, 2026 | August 28, 2026 |
| Context window | 1.05M | 1.05M |
| Input $/Mtok | $1.358 | $0.834 |
| Output $/Mtok | $2.976 | $2.501 |
| Coding index | 38.7 | 44.3 |
| Independent measurements behind it | 5 | 3 |
| Cost per solved task | $0.57 | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is DeepSeek V4 Pro 0423 better than Hy4 preview for coding?
Pick Hy4 preview for coding: it scores 44.3 on the Vibe Coding Index against 38.7 for DeepSeek V4 Pro 0423, though on 3 benchmarks against 5. Pick DeepSeek V4 Pro 0423 for work measured by Terminal-Bench 4 (Vals AI), where it leads.
Which is cheaper, DeepSeek V4 Pro 0423 or Hy4 preview?
DeepSeek V4 Pro 0423 costs $1.358 per million input tokens and $2.976 per million output tokens; Hy4 preview costs $0.834 and $2.501. Cost per solved task is known for DeepSeek V4 Pro 0423 only.
Which has the bigger context window, DeepSeek V4 Pro 0423 or Hy4 preview?
Neither: both take 1.05M tokens.
How many benchmarks is the DeepSeek V4 Pro 0423 vs Hy4 preview comparison based on?
17 benchmarks have a result for both models as of October 22, 2026. DeepSeek V4 Pro 0423's coding index rests on 5 independent benchmarks and Hy4 preview's on 3. Where only one model has a result, the page says so instead of naming a winner.