GPT-5.2 vs Llama 3.3 70B Instruct
GPT-5.2 leads Llama 3.3 70B Instruct on the Artificial Analysis Intelligence Index, 30.4 to 7.7, and Llama 3.3 70B Instruct costs 19.4x less per output token ($0.72 vs $14.00 per million). Updated October 21, 2026; 46 benchmarks have a result for both.
GPT-5.2 or Llama 3.3 70B Instruct: which to pick
Only GPT-5.2 has a coding score, 55.9 on the Vibe Coding Index; Llama 3.3 70B Instruct has none, so the two cannot be ranked on coding.
Pick Llama 3.3 70B Instruct for output tokens 19.4x cheaper ($0.72 vs $14.00 per million).
- Shared results
- 46
- GPT-5.2 only
- 65
- Llama 3.3 70B Instruct only
- 5
- 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-5.2 has a coding score, so no call is made.
One side not measured
Intelligence
AA Intelligence Index
GPT-5.2
GPT-5.2 leads, 30.4 against 7.7.
Same board, both measured
Reasoning
GPQA Diamond
No call
Only GPT-5.2 has a GPQA Diamond result, 91.7%.
Directional only
Math
AIME
No call
Only GPT-5.2 has a AIME result, 96.9%.
Directional only
Knowledge
MMLU Pro
No call
Only GPT-5.2 has a MMLU Pro result, 86.2%.
Directional only
Web design
Design Arena: websites
No call
Only GPT-5.2 has a Design Arena: websites result, 1195.
Directional only
Long documents
Listed context window
GPT-5.2
GPT-5.2 takes 400K tokens against 131K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Llama 3.3 70B Instruct
GPT-5.2 $0.0088, Llama 3.3 70B Instruct $0.0008 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only GPT-5.2 has the number, $1.91.
Directional only
At a glance
- Coding index
- GPT-5.2 55.9 from 3 benchmarks, Llama 3.3 70B Instruct not benchmarked.
- Widest category
- Math, 66.3 points apart on average over 5 shared benchmarks.
- Widest single gap
- AIME 2025 (Artificial Analysis run): GPT-5.2 99.0% against Llama 3.3 70B Instruct 7.7%, 91.3 points apart on a 0 to 100 scale.
- Output price
- GPT-5.2 $14.00 against Llama 3.3 70B Instruct $0.72 per million output tokens, 19.4x apart.
- Context window
- GPT-5.2 takes 400K tokens, Llama 3.3 70B Instruct 131K.
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-5.2 | Bars | Llama 3.3 70B Instruct | Ties |
|---|---|---|---|---|
| Coding | 5 | 0 | ||
| Intelligence | 2 | 0 | ||
| Agentic work | 1 | 0 | ||
| Reasoning | 7 | 0 | ||
| Math | 5 | 0 | ||
| Knowledge | 5 | 0 | ||
| General | 21 | 0 | ||
| All shared boards | 46 | 0 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index50
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| SWE-bench Verified | 72.8% | Not measured | Only GPT-5.2 measured |
| SWE-bench Multilingual | 66.7% | Not measured | Only GPT-5.2 measured |
| LMArena WebDev | 1417 | Not measured | Only GPT-5.2 measured |
| AA Coding Index | Not measured | 11.9 | Only Llama 3.3 70B Instruct measured |
| LiveCodeBench (Vals AI) | 85.4% | Not measured | Only GPT-5.2 measured |
| SWE-bench Verified (Vals AI) | 75.8% | Not measured | Only GPT-5.2 measured |
| Vibe Code Bench (Vals AI) | 53.5% | Not measured | Only GPT-5.2 measured |
| LiveCodeBench (Artificial Analysis run) | 89.4% | 28.8% | GPT-5.2 |
| Terminal-Bench Hard (Artificial Analysis run) | 47.0% | 3.0% | GPT-5.2 |
| Terminal-Bench | 64.9% | Not measured | Only GPT-5.2 measured |
| GSO-Bench | 27.5% | Not measured | Only GPT-5.2 measured |
| WeirdML | 72.2% | 14.4% | GPT-5.2 |
| ALE-Bench | 1293.5 | Not measured | Only GPT-5.2 measured |
| LMArena Text: coding | 1442 | 1268 | GPT-5.2 |
| LMArena Text: software and IT | 1434 | 1273 | GPT-5.2 |
| LMArena WebDev: HTML | 1427 | Not measured | Only GPT-5.2 measured |
| Neither model has a result on the other 22 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| AA Intelligence Index | 30.4 | 7.7 | GPT-5.2 |
| Epoch Capabilities Index | 153.7 | 127.3 | GPT-5.2 |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index10
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| Design Arena: agents building web apps | 1086 | Not measured | Only GPT-5.2 measured |
| Design Arena: agents building full-stack apps | 1043 | Not measured | Only GPT-5.2 measured |
| Design Arena: agents building mobile apps | 1093 | Not measured | Only GPT-5.2 measured |
| Design Arena: agents building Android apps | 1058 | Not measured | Only GPT-5.2 measured |
| Design Arena: agents building Godot games | 1142 | Not measured | Only GPT-5.2 measured |
| τ²-Bench Telecom | 84.8% | 26.6% | GPT-5.2 |
| METR time horizons | 75.3% | Not measured | Only GPT-5.2 measured |
| Remote Labor Index | 2.5% | Not measured | Only GPT-5.2 measured |
| GDPval | 49.7% | Not measured | Only GPT-5.2 measured |
| BALROG | Not measured | 23.0% | Only Llama 3.3 70B Instruct measured |
| DeepResearch Bench | 41.1% | Not measured | Only GPT-5.2 measured |
| Vending-Bench 2 | $3,591 | Not measured | Only GPT-5.2 measured |
| Neither model has a result on the other 27 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index70
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| GPQA Diamond | 91.7% | Not measured | Only GPT-5.2 measured |
| MMMU Pro (Vals AI) | 86.7% | Not measured | Only GPT-5.2 measured |
| GPQA Diamond (Artificial Analysis run) | 90.3% | 49.8% | GPT-5.2 |
| Humanity's Last Exam (Artificial Analysis run) | 37.7% | 3.6% | GPT-5.2 |
| GPQA Diamond (Epoch AI run) | 91.4% | 47.4% | GPT-5.2 |
| ARC-AGI | 86.2% | Not measured | Only GPT-5.2 measured |
| ARC-AGI-2 | 52.9% | Not measured | Only GPT-5.2 measured |
| SimpleBench | 45.8% | 19.9% | GPT-5.2 |
| Chess puzzles | 48.0% | Not measured | Only GPT-5.2 measured |
| Mystery game puzzles | 23.0% | Not measured | Only GPT-5.2 measured |
| EBR-bench | 23.0% | Not measured | Only GPT-5.2 measured |
| VPCT | 84.0% | Not measured | Only GPT-5.2 measured |
| LMArena Text: hard prompts | 1428 | 1257 | GPT-5.2 |
| LMArena Text: expert | 1445 | 1225 | GPT-5.2 |
| LMArena Text: hard prompts in English | 1431 | 1277 | GPT-5.2 |
| LMArena Vision: diagrams | 1265 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: homework | 1276 | Not measured | Only GPT-5.2 measured |
| Neither model has a result on the other 4 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index50
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| AIME | 96.9% | Not measured | Only GPT-5.2 measured |
| MGSM (Vals AI) | 94.0% | Not measured | Only GPT-5.2 measured |
| AA Math Index | 99.0 | 7.7 | GPT-5.2 |
| AIME 2025 (Artificial Analysis run) | 99.0% | 7.7% | GPT-5.2 |
| OTIS Mock AIME | 96.1% | 5.1% | GPT-5.2 |
| FrontierMath, tiers 1 to 3 | 67.4% | Not measured | Only GPT-5.2 measured |
| FrontierMath, tier 4 | 31.7% | Not measured | Only GPT-5.2 measured |
| ProofBench | 15.0% | Not measured | Only GPT-5.2 measured |
| MATH level 5 | Not measured | 41.6% | Only Llama 3.3 70B Instruct measured |
| LMArena Text: math | 1440 | 1267 | GPT-5.2 |
| LMArena Text: mathematical work | 1442 | 1267 | GPT-5.2 |
| Neither model has a result on the other 2 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index50
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| MMLU Pro | 86.2% | Not measured | Only GPT-5.2 measured |
| LegalBench (Vals AI) | 82.8% | Not measured | Only GPT-5.2 measured |
| CaseLaw (Vals AI) | 66.0% | Not measured | Only GPT-5.2 measured |
| CorpFin (Vals AI) | 65.9% | Not measured | Only GPT-5.2 measured |
| TaxEval (Vals AI) | 75.8% | Not measured | Only GPT-5.2 measured |
| MortgageTax (Vals AI) | 67.1% | Not measured | Only GPT-5.2 measured |
| MedQA (Vals AI) | 94.1% | Not measured | Only GPT-5.2 measured |
| MedCode (Vals AI) | 49.8% | Not measured | Only GPT-5.2 measured |
| MMLU Pro (Artificial Analysis run) | 87.4% | 71.3% | GPT-5.2 |
| SimpleQA Verified | 37.1% | Not measured | Only GPT-5.2 measured |
| MMLU | Not measured | 86.3% | Only Llama 3.3 70B Instruct measured |
| LMArena Text: science | 1425 | 1284 | GPT-5.2 |
| LMArena Text: medicine and healthcare | 1425 | 1275 | GPT-5.2 |
| LMArena Text: legal and government | 1427 | 1284 | GPT-5.2 |
| LMArena Text: business and finance | 1414 | 1243 | GPT-5.2 |
| Neither model has a result on the other 2 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index8 boards, none shared
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| Design Arena: websites | 1195 | Not measured | Only GPT-5.2 measured |
| Design Arena: UI components | 1190 | Not measured | Only GPT-5.2 measured |
| Design Arena: game dev | 1195 | Not measured | Only GPT-5.2 measured |
| Design Arena: data viz | 1202 | Not measured | Only GPT-5.2 measured |
| Design Arena: 3D | 1079 | Not measured | Only GPT-5.2 measured |
| Design Arena: all code categories | 1170 | Not measured | Only GPT-5.2 measured |
| Design Arena: SVG | 1139 | Not measured | Only GPT-5.2 measured |
| Design Arena: ASCII art | 1152 | Not measured | Only GPT-5.2 measured |
| Neither model has a result on the other 1 board here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index210
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | Leader |
|---|---|---|---|
| SAGE (Vals AI) | 49.3% | Not measured | Only GPT-5.2 measured |
| MedScribe (Vals AI) | 84.4% | Not measured | Only GPT-5.2 measured |
| IFBench | 75.4% | 47.1% | GPT-5.2 |
| Long context reasoning (AA-LCR) | 82.7% | 15.7% | GPT-5.2 |
| DTBench | 90.9% | 59.5% | GPT-5.2 |
| LMCA | 43.9% | 17.5% | GPT-5.2 |
| Fiction.LiveBench | Not measured | 33.3% | Only Llama 3.3 70B Instruct measured |
| CL-bench | 18.2% | Not measured | Only GPT-5.2 measured |
| Furniture assembly | 38.3% | Not measured | Only GPT-5.2 measured |
| LMArena Text | 1416 | 1274 | GPT-5.2 |
| LMArena Text: instruction following | 1409 | 1242 | GPT-5.2 |
| LMArena Text: creative writing | 1376 | 1250 | GPT-5.2 |
| LMArena Vision | 1246 | Not measured | Only GPT-5.2 measured |
| LMArena Text: English | 1422 | 1303 | GPT-5.2 |
| LMArena Text: non-English | 1405 | 1236 | GPT-5.2 |
| LMArena Text: Chinese | 1447 | 1217 | GPT-5.2 |
| LMArena Text: German | 1444 | 1251 | GPT-5.2 |
| LMArena Text: Japanese | 1420 | 1150 | GPT-5.2 |
| LMArena Text: Korean | 1374 | 1142 | GPT-5.2 |
| LMArena Text: French | 1454 | 1282 | GPT-5.2 |
| LMArena Text: Spanish | 1413 | 1269 | GPT-5.2 |
| LMArena Text: Russian | 1415 | 1252 | GPT-5.2 |
| LMArena Text: Polish | 1426 | 1259 | GPT-5.2 |
| LMArena Text: multi-turn | 1419 | 1280 | GPT-5.2 |
| LMArena Text: longer queries | 1413 | 1255 | GPT-5.2 |
| LMArena Text: writing and language | 1392 | 1255 | GPT-5.2 |
| LMArena Text: entertainment and media | 1378 | 1241 | GPT-5.2 |
| LMArena Vision: English | 1243 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: Chinese | 1305 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: reading text | 1257 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: humor | 1227 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: captioning | 1267 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: recognizing things | 1232 | Not measured | Only GPT-5.2 measured |
| LMArena Vision: creative writing | 1225 | Not measured | Only GPT-5.2 measured |
| LMArena Document | 1405 | Not measured | Only GPT-5.2 measured |
| Neither model has a result on the other 11 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | GPT-5.2 | Llama 3.3 70B Instruct | 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
Llama 3.3 70B Instruct costs less
Repository review
50K input + 3K output tokens
Llama 3.3 70B Instruct costs less
Agent loop
200K input + 10K output tokens
Llama 3.3 70B Instruct 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-5.2 | Meta Llama 3.3 70B Instruct |
|---|---|---|
| Maker | OpenAI | Meta Llama |
| Weights | No public weights listed | Open weights |
| Released | December 10, 2025 | December 6, 2024 |
| Context window | 400K | 131K |
| Input $/Mtok | $1.75 | $0.45 |
| Output $/Mtok | $14.00 | $0.72 |
| Coding index | 55.9 | Not benchmarked |
| Independent measurements behind it | 3 | 0 |
| Cost per solved task | $1.91 | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is GPT-5.2 better than Llama 3.3 70B Instruct for coding?
Only GPT-5.2 has a coding score, 55.9 on the Vibe Coding Index; Llama 3.3 70B Instruct has none, so the two cannot be ranked on coding. Pick Llama 3.3 70B Instruct for output tokens 19.4x cheaper ($0.72 vs $14.00 per million).
Which is cheaper, GPT-5.2 or Llama 3.3 70B Instruct?
GPT-5.2 costs $1.75 per million input tokens and $14.00 per million output tokens; Llama 3.3 70B Instruct costs $0.45 and $0.72. Cost per solved task is known for GPT-5.2 only.
Which has the bigger context window, GPT-5.2 or Llama 3.3 70B Instruct?
GPT-5.2, with 400K tokens against 131K.
How many benchmarks is the GPT-5.2 vs Llama 3.3 70B Instruct comparison based on?
46 benchmarks have a result for both models as of October 21, 2026. GPT-5.2's coding index rests on 3 independent benchmarks and Llama 3.3 70B Instruct's on 0. Where only one model has a result, the page says so instead of naming a winner.