# GPT-5.4 vs Llama 3-8B

> GPT-5.4 is the stronger model overall, scoring 59.5 to 25.3 on the Noometry Index. Llama 3-8B costs 40× less per token, which makes it the better buy when GPT-5.4's lead doesn't matter for your workload.

- Canonical page: https://noometry.com/compare/gpt-5-4-vs-llama-3-8b
- Last updated: 2026-10-10
- Shared benchmarks: 23

## Summary

- They share 23 benchmarks with published results for both. GPT-5.4 scores higher in 8 categories and Llama 3-8B in 0 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where GPT-5.4 leads 73.3 to 8.3.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 97.8% for GPT-5.4 and 1.9% for Llama 3-8B.
- Llama 3-8B is cheaper at $0.14 / $0.14 per million input/output tokens, against $2.50 / $15 for GPT-5.4.
- GPT-5.4 accepts more context: 1.05M tokens versus 8K.
- Llama 3-8B has downloadable open weights; the other is API-only.

## Snapshot

| | GPT-5.4 | Llama 3-8B |
|---|---|---|
| Provider | OpenAI | Meta |
| Noometry Index | 59.5 | 25.3 |
| Rank | 14 | 338 |
| Context | 1.05M | 8K |
| Input $/M | $2.50 | $0.14 |
| Output $/M | $15 | $0.14 |
| Weights | Proprietary | Open |

## Coding

- GPT-5.4: 52.8 (#28)
- Llama 3-8B: 30.5 (#285)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| LMArena Coding | 1497 | 1152 |
| SWE-bench Verified | 76.9% | — |
| DeepSWE | 51.8% | — |
| LMArena WebDev | 1465 | — |
| SciCode | 56.6% | — |
| GSO | 31.4% | — |
| WeirdML | 77.7% | — |
| BigCodeBench Instruct | — | 31.9% |
| MirrorCode | 15.6% | — |
| BigCodeBench Complete | — | 36.9% |
| ALE-Bench | 1,607 | — |
| AlgoTune | 1.85 | — |
| HumanEval+ | — | 56.7% |
| MBPP+ | — | 54.8% |

## Agentic & Tool Use

- GPT-5.4: 46.5 (#12)
- Llama 3-8B: —

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| Terminal-Bench | 81.8% | — |
| APEX-Agents | 52.4% | — |
| τ²-bench Banking | 39.4% | — |
| DeepResearch Bench | 35.1% | — |
| PostTrainBench | 19% | — |
| GBAEval | 45.1% | — |
| LMArena Search | 1197 | — |
| METR Time Horizons | 74.3% | — |
| Vending-Bench 2 | 6,144 | — |

## Reasoning

- GPT-5.4: 61.7 (#17)
- Llama 3-8B: 13.8 (#321)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| Chess Puzzles | 44% | 0% |
| LMArena Hard Prompts | 1485 | 1133 |
| DTBench | 94.4% | 43.9% |
| Epoch Capabilities Index | 156.92 | 116.33 |
| ForecastBench | 59.5 | 58.6 |
| ARC-AGI-2 | 74% | — |
| Kagi LLM Benchmark | 63.8% | — |
| NYT Connections (extended) | 91.3% | — |
| ARC-AGI-1 | 93.7% | — |
| CritPt | 23.4% | — |
| EnigmaEval | 16% | — |
| Thematic Generalization | 80% | — |
| EBR-Bench | 25.4% | — |
| Mystery Game Puzzles | 37% | — |
| LMCA | 52% | — |
| Adversarial NLI | — | 57.3% |
| WinoGrande | — | 75.7% |

## Math

- GPT-5.4: 73.3 (#14)
- Llama 3-8B: 8.3 (#317)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 97.8% | 1.9% |
| LMArena Math | 1488 | 1151 |
| FrontierMath (Tiers 1-3) | 78.6% | — |
| FrontierMath Tier 4 | 49% | — |
| MathArena Final-Answer Competitions | 83.1% | — |
| ProofBench | 56% | — |
| MATH Level 5 | — | 6.1% |
| FrontierMath (Feb 2025 set) | 47.6% | — |
| FrontierMath Tier 4 (v1) | 27.1% | — |

## Knowledge

- GPT-5.4: 65.3 (#13)
- Llama 3-8B: 7.8 (#303)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| GPQA Diamond | 93.3% | 26.1% |
| LMArena Expert | 1507 | 1113 |
| Humanity's Last Exam | 36.2% | — |
| SimpleQA Verified | 45.1% | — |
| Vectara Hallucination Rate | 7% | — |
| ARC (AI2) Challenge | — | 82.8% |
| MMLU | — | 68.8% |
| OpenBookQA | — | 82.6% |
| TriviaQA | — | 67.7% |

## Multimodal

- GPT-5.4: 43.8 (#19)
- Llama 3-8B: —

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| LMArena Vision | 1302 | — |
| Blueprint-Bench 2 | 27.1% | — |
| Furniture Assembly | 37.5% | — |
| LMArena Document | 1471 | — |

## Multilingual

- GPT-5.4: 56.4 (#22)
- Llama 3-8B: 31.0 (#256)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| LMArena Non-English | 1465 | 1098 |
| LMArena Chinese | 1520 | 1076 |
| LMArena French | 1493 | 1158 |
| LMArena German | 1472 | 1104 |
| LMArena Japanese | 1484 | 967 |
| LMArena Korean | 1449 | 1003 |
| LMArena Russian | 1480 | 1109 |
| LMArena Spanish | 1453 | 1172 |

## Instruction Following

- GPT-5.4: 77.3 (#25)
- Llama 3-8B: 58.7 (#255)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| LMArena Instruction Following | 1469 | 1127 |

## Long Context

- GPT-5.4: 50.3 (#8)
- Llama 3-8B: 34.2 (#246)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| LMArena Longer Query | 1474 | 1128 |
| CL-bench | 27.9% | — |
| CL-bench Life | 21.7% | — |

## Writing & Preference

- GPT-5.4: 72.1 (#18)
- Llama 3-8B: 37.7 (#251)

| Benchmark | GPT-5.4 | Llama 3-8B |
|---|---|---|
| LMArena Text | 1469 | 1166 |
| LMArena Creative Writing | 1439 | 1150 |
| LMArena Multi-Turn | 1482 | 1152 |
| EQ-Bench Creative Writing | 1840 | — |
| EQ-Bench 4 | 1272 | — |

## FAQ

### Is GPT-5.4 better than Llama 3-8B?

GPT-5.4 is the stronger model overall, scoring 59.5 to 25.3 on the Noometry Index. Llama 3-8B costs 40× less per token, which makes it the better buy when GPT-5.4's lead doesn't matter for your workload.

### Which is cheaper, GPT-5.4 or Llama 3-8B?

Llama 3-8B is cheaper. It lists at $0.14 per million input tokens and $0.14 per million output tokens; GPT-5.4 lists at $2.50 and $15.

### Is GPT-5.4 or Llama 3-8B better for coding?

GPT-5.4 scores higher on coding benchmarks: 52.8 versus 30.5 in the Noometry coding category.

### Which has the bigger context window?

GPT-5.4 does, with 1.05M tokens against 8K.

### How many benchmarks do GPT-5.4 and Llama 3-8B share?

23 benchmarks have published results for both models. GPT-5.4 has 68 scored results on Noometry and Llama 3-8B has 34.