GPT-5.4 vs Hy3 preview
GPT-5.4 leads Hy3 preview on the Artificial Analysis Intelligence Index, 39.0 to 25.3, and Hy3 preview costs 25.0x less per output token ($0.60 vs $15.00 per million). Updated October 21, 2026; 35 benchmarks have a result for both.
GPT-5.4 or Hy3 preview: which to pick
Only GPT-5.4 has a coding score, 64.4 on the Vibe Coding Index; Hy3 preview has none, so the two cannot be ranked on coding.
Pick Hy3 preview for output tokens 25.0x cheaper ($0.60 vs $15.00 per million) and work measured by τ²-Bench Telecom, where it leads.
- Shared results
- 35
- GPT-5.4 only
- 75
- Hy3 preview only
- 1
- Like-for-like categories
- 7 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
No call
Only GPT-5.4 has a coding score, so no call is made.
One side not measured
Intelligence
AA Intelligence Index
GPT-5.4
GPT-5.4 leads, 39.0 against 25.3.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only Hy3 preview has a AA Agentic Index result, 24.1.
Directional only
Reasoning
GPQA Diamond
No call
Only GPT-5.4 has a GPQA Diamond result, 91.7%.
Directional only
Math
AIME
No call
Only GPT-5.4 has a AIME result, 96.7%.
Directional only
Knowledge
MMLU Pro
No call
Only GPT-5.4 has a MMLU Pro result, 87.5%.
Directional only
Web design
Design Arena: websites
No call
Only GPT-5.4 has a Design Arena: websites result, 1199.
Directional only
Long documents
Listed context window
GPT-5.4
GPT-5.4 takes 1.05M tokens against 262K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Hy3 preview
GPT-5.4 $0.010, Hy3 preview $0.0005 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only GPT-5.4 has the number, $1.02.
Directional only
At a glance
- Coding index
- GPT-5.4 64.4 from 3 benchmarks, Hy3 preview not benchmarked.
- Widest category
- Coding, 16.1 points apart on average over 7 shared benchmarks.
- Widest single gap
- Terminal-Bench Hard (Artificial Analysis run): GPT-5.4 57.6% against Hy3 preview 34.1%, 23.5 points apart on a 0 to 100 scale.
- Output price
- GPT-5.4 $15.00 against Hy3 preview $0.60 per million output tokens, 25.0x apart.
- Context window
- GPT-5.4 takes 1.05M tokens, Hy3 preview 262K.
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 | GPT-5.4 | Bars | Hy3 preview | Ties |
|---|---|---|---|---|
| Coding | 7 | 0 | ||
| Intelligence | 1 | 0 | ||
| Agentic work | 0 | 1 | ||
| Reasoning | 5 | 0 | ||
| Math | 2 | 0 | ||
| Knowledge | 4 | 0 | ||
| General | 15 | 0 | ||
| All shared boards | 34 | 1 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index70
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| LMArena WebDev | 1464 | 1357 | GPT-5.4 |
| AA Coding Index | 71.1 | 58.8 | GPT-5.4 |
| LiveCodeBench (Vals AI) | 84.1% | Not measured | Only GPT-5.4 measured |
| SWE-bench Verified (Vals AI) | 78.2% | Not measured | Only GPT-5.4 measured |
| Vibe Code Bench (Vals AI) | 67.4% | Not measured | Only GPT-5.4 measured |
| Code migration (Vals AI) | 35.0% | Not measured | Only GPT-5.4 measured |
| ProgramBench (Vals AI) | 0.5% | Not measured | Only GPT-5.4 measured |
| SkillsBench (Vals AI) | 51.7% | Not measured | Only GPT-5.4 measured |
| Terminal-Bench Hard (Artificial Analysis run) | 57.6% | 34.1% | GPT-5.4 |
| Terminal-Bench | 81.8% | Not measured | Only GPT-5.4 measured |
| GSO-Bench | 31.4% | Not measured | Only GPT-5.4 measured |
| WeirdML | 77.7% | Not measured | Only GPT-5.4 measured |
| MirrorCode | 15.6% | Not measured | Only GPT-5.4 measured |
| ALE-Bench | 1607.0 | Not measured | Only GPT-5.4 measured |
| LMArena Text: coding | 1496 | 1421 | GPT-5.4 |
| LMArena Text: software and IT | 1491 | 1423 | GPT-5.4 |
| LMArena WebDev: React | 1454 | 1349 | GPT-5.4 |
| LMArena WebDev: HTML | 1478 | 1356 | GPT-5.4 |
| LMArena WebDev: from an image | 1436 | Not measured | Only GPT-5.4 measured |
| Neither model has a result on the other 20 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index10
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| AA Intelligence Index | 39.0 | 25.3 | GPT-5.4 |
| Epoch Capabilities Index | 157.3 | Not measured | Only GPT-5.4 measured |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| AA Agentic Index | Not measured | 24.1 | Only Hy3 preview measured |
| Design Arena: agents building web apps | 1059 | Not measured | Only GPT-5.4 measured |
| Design Arena: agents building full-stack apps | 1017 | Not measured | Only GPT-5.4 measured |
| Design Arena: agents building mobile apps | 1087 | Not measured | Only GPT-5.4 measured |
| Design Arena: agents building Android apps | 1030 | Not measured | Only GPT-5.4 measured |
| Design Arena: agents building Godot games | 1135 | Not measured | Only GPT-5.4 measured |
| τ²-Bench Telecom | 87.1% | 92.7% | Hy3 preview |
| APEX-Agents | 52.4% | Not measured | Only GPT-5.4 measured |
| METR time horizons | 74.3% | Not measured | Only GPT-5.4 measured |
| DeepResearch Bench | 35.1% | Not measured | Only GPT-5.4 measured |
| PostTrainBench | 19.0% | Not measured | Only GPT-5.4 measured |
| Vending-Bench 2 | $6,144 | Not measured | Only GPT-5.4 measured |
| Neither model has a result on the other 29 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index50
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| GPQA Diamond | 91.7% | Not measured | Only GPT-5.4 measured |
| MMMU Pro (Vals AI) | 87.5% | Not measured | Only GPT-5.4 measured |
| GPQA Diamond (Artificial Analysis run) | 92.0% | 86.7% | GPT-5.4 |
| Humanity's Last Exam (Artificial Analysis run) | 43.7% | 27.8% | GPT-5.4 |
| GPQA Diamond (Epoch AI run) | 93.3% | Not measured | Only GPT-5.4 measured |
| ARC-AGI | 93.7% | Not measured | Only GPT-5.4 measured |
| ARC-AGI-2 | 74.0% | Not measured | Only GPT-5.4 measured |
| Chess puzzles | 48.0% | Not measured | Only GPT-5.4 measured |
| Mystery game puzzles | 37.0% | Not measured | Only GPT-5.4 measured |
| EBR-bench | 25.4% | Not measured | Only GPT-5.4 measured |
| LMArena Text: hard prompts | 1485 | 1413 | GPT-5.4 |
| LMArena Text: expert | 1507 | 1422 | GPT-5.4 |
| LMArena Text: hard prompts in English | 1479 | 1416 | GPT-5.4 |
| LMArena Vision: diagrams | 1327 | Not measured | Only GPT-5.4 measured |
| LMArena Vision: homework | 1336 | Not measured | Only GPT-5.4 measured |
| Kagi LLM Benchmark | 63.8% | Not measured | Only GPT-5.4 measured |
| Neither model has a result on the other 6 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| AIME | 96.7% | Not measured | Only GPT-5.4 measured |
| OTIS Mock AIME | 97.8% | Not measured | Only GPT-5.4 measured |
| FrontierMath, tiers 1 to 3 | 78.6% | Not measured | Only GPT-5.4 measured |
| FrontierMath, tier 4 | 49.0% | Not measured | Only GPT-5.4 measured |
| ProofBench | 56.0% | Not measured | Only GPT-5.4 measured |
| LMArena Text: math | 1488 | 1424 | GPT-5.4 |
| LMArena Text: mathematical work | 1497 | 1442 | GPT-5.4 |
| Neither model has a result on the other 6 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index40
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| MMLU Pro | 87.5% | Not measured | Only GPT-5.4 measured |
| LegalBench (Vals AI) | 86.0% | Not measured | Only GPT-5.4 measured |
| CaseLaw (Vals AI) | 63.8% | Not measured | Only GPT-5.4 measured |
| CorpFin (Vals AI) | 65.3% | Not measured | Only GPT-5.4 measured |
| TaxEval (Vals AI) | 74.0% | Not measured | Only GPT-5.4 measured |
| MortgageTax (Vals AI) | 68.3% | Not measured | Only GPT-5.4 measured |
| MedQA (Vals AI) | 96.1% | Not measured | Only GPT-5.4 measured |
| MedCode (Vals AI) | 41.3% | Not measured | Only GPT-5.4 measured |
| SimpleQA Verified | 45.1% | Not measured | Only GPT-5.4 measured |
| LMArena Text: science | 1479 | 1416 | GPT-5.4 |
| LMArena Text: medicine and healthcare | 1458 | 1405 | GPT-5.4 |
| LMArena Text: legal and government | 1486 | 1430 | GPT-5.4 |
| LMArena Text: business and finance | 1474 | 1397 | GPT-5.4 |
| Neither model has a result on the other 5 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index8 boards, none shared
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| Design Arena: websites | 1199 | Not measured | Only GPT-5.4 measured |
| Design Arena: UI components | 1215 | Not measured | Only GPT-5.4 measured |
| Design Arena: game dev | 1195 | Not measured | Only GPT-5.4 measured |
| Design Arena: data viz | 1196 | Not measured | Only GPT-5.4 measured |
| Design Arena: 3D | 1066 | Not measured | Only GPT-5.4 measured |
| Design Arena: all code categories | 1180 | Not measured | Only GPT-5.4 measured |
| Design Arena: SVG | 1142 | Not measured | Only GPT-5.4 measured |
| Design Arena: ASCII art | 1188 | Not measured | Only GPT-5.4 measured |
| Neither model has a result on the other 1 board here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index150
| Benchmark | GPT-5.4 | Hy3 preview | Leader |
|---|---|---|---|
| SAGE (Vals AI) | 43.3% | Not measured | Only GPT-5.4 measured |
| MedScribe (Vals AI) | 77.5% | Not measured | Only GPT-5.4 measured |
| IFBench | 73.9% | 63.1% | GPT-5.4 |
| Long context reasoning (AA-LCR) | 82.0% | 64.7% | GPT-5.4 |
| DTBench | 94.4% | Not measured | Only GPT-5.4 measured |
| LMCA | 52.0% | Not measured | Only GPT-5.4 measured |
| CL-bench | 27.9% | Not measured | Only GPT-5.4 measured |
| Furniture assembly | 37.5% | Not measured | Only GPT-5.4 measured |
| CL-bench Life | 21.7% | Not measured | Only GPT-5.4 measured |
| LMArena Text | 1469 | 1401 | GPT-5.4 |
| LMArena Text: instruction following | 1469 | 1383 | GPT-5.4 |
| LMArena Text: creative writing | 1439 | 1339 | GPT-5.4 |
| LMArena Vision | 1303 | Not measured | Only GPT-5.4 measured |
| LMArena Text: English | 1467 | 1409 | GPT-5.4 |
| LMArena Text: non-English | 1465 | 1390 | GPT-5.4 |
| LMArena Text: Chinese | 1520 | 1435 | GPT-5.4 |
| LMArena Text: German | 1472 | Not measured | Only GPT-5.4 measured |
| LMArena Text: Japanese | 1485 | Not measured | Only GPT-5.4 measured |
| LMArena Text: Korean | 1449 | Not measured | Only GPT-5.4 measured |
| LMArena Text: French | 1493 | 1425 | GPT-5.4 |
| LMArena Text: Spanish | 1452 | 1399 | GPT-5.4 |
| LMArena Text: Russian | 1480 | 1399 | GPT-5.4 |
| LMArena Text: Polish | 1472 | Not measured | Only GPT-5.4 measured |
| LMArena Text: multi-turn | 1482 | 1411 | GPT-5.4 |
| LMArena Text: longer queries | 1473 | 1402 | GPT-5.4 |
| LMArena Text: writing and language | 1458 | 1370 | GPT-5.4 |
| LMArena Text: entertainment and media | 1438 | 1352 | GPT-5.4 |
| LMArena Vision: English | 1301 | Not measured | Only GPT-5.4 measured |
| LMArena Vision: Chinese | 1354 | Not measured | Only GPT-5.4 measured |
| LMArena Vision: reading text | 1312 | Not measured | Only GPT-5.4 measured |
| LMArena Vision: humor | 1284 | Not measured | Only GPT-5.4 measured |
| LMArena Vision: recognizing things | 1251 | Not measured | Only GPT-5.4 measured |
| LMArena Vision: creative writing | 1283 | Not measured | Only GPT-5.4 measured |
| LMArena Document | 1471 | Not measured | Only GPT-5.4 measured |
| Neither model has a result on the other 12 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | GPT-5.4 | Hy3 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
Hy3 preview costs less
Repository review
50K input + 3K output tokens
Hy3 preview costs less
Agent loop
200K input + 10K output tokens
Hy3 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 | OpenAI GPT-5.4 | Tencent Hy3 preview |
|---|---|---|
| Maker | OpenAI | Tencent |
| Weights | No public weights listed | Open weights |
| Released | March 5, 2026 | April 22, 2026 |
| Context window | 1.05M | 262K |
| Input $/Mtok | $2.50 | $0.18 |
| Output $/Mtok | $15.00 | $0.60 |
| Coding index | 64.4 | Not benchmarked |
| Independent measurements behind it | 3 | 1 |
| Cost per solved task | $1.02 | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is GPT-5.4 better than Hy3 preview for coding?
Only GPT-5.4 has a coding score, 64.4 on the Vibe Coding Index; Hy3 preview has none, so the two cannot be ranked on coding. Pick Hy3 preview for output tokens 25.0x cheaper ($0.60 vs $15.00 per million) and work measured by τ²-Bench Telecom, where it leads.
Which is cheaper, GPT-5.4 or Hy3 preview?
GPT-5.4 costs $2.50 per million input tokens and $15.00 per million output tokens; Hy3 preview costs $0.18 and $0.60. Cost per solved task is known for GPT-5.4 only.
Which has the bigger context window, GPT-5.4 or Hy3 preview?
GPT-5.4, with 1.05M tokens against 262K.
How many benchmarks is the GPT-5.4 vs Hy3 preview comparison based on?
35 benchmarks have a result for both models as of October 21, 2026. GPT-5.4's coding index rests on 3 independent benchmarks and Hy3 preview's on 1. Where only one model has a result, the page says so instead of naming a winner.