Updated March 26, 2026· Based on independent benchmark data
GPT-5.3 Codex (xhigh) leads in intelligence with a score of 54.0 vs 11.9. Llama 3.2 Instruct 90B (Vision) is 2.4x cheaper at $0.72/1M tokens vs $1.75/1M. For speed, GPT-5.3 Codex (xhigh) wins at 72 tok/s vs 42 tok/s.
| Metric | Llama 3.2 Instruct 90B (Vision) | GPT-5.3 Codex (xhigh) |
|---|---|---|
| Intelligence Score | 11.9 | 54.0 |
| Coding Score | N/A | 53.1 |
| Math Score | N/A | N/A |
| Speed (tok/s) | 42 tok/s | 72 tok/s |
| Latency (TTFT) | 0.38s | 83.91s |
| Input Price / 1M tokens | $0.72 | $1.75 |
| Output Price / 1M tokens | $0.72 | $14 |
| Context Window | N/A |
GPT-5.3 Codex (xhigh) outperforms Llama 3.2 Instruct 90B (Vision) on the intelligence index with a score of 54.0 compared to 11.9.
GPT-5.3 Codex (xhigh) generates output significantly faster at 72 tok/s compared to Llama 3.2 Instruct 90B (Vision)'s 42 tok/s, making it 1.7x faster for streaming responses. Time to first token is 0.38s for Llama 3.2 Instruct 90B (Vision) vs 83.91s for GPT-5.3 Codex (xhigh), which affects perceived responsiveness in interactive applications.
Llama 3.2 Instruct 90B (Vision) is more affordable at $0.72/1M input tokens ($0.72/1M output), while GPT-5.3 Codex (xhigh) costs $1.75/1M input ($14/1M output). That makes GPT-5.3 Codex (xhigh) 2.4x more expensive per token, which can add up significantly at scale. For a typical workload of 100 requests per day at 2,000 tokens each, Llama 3.2 Instruct 90B (Vision) would cost approximately $4.32/month vs $10.50/month for GPT-5.3 Codex (xhigh) in input costs alone.
Choose Llama 3.2 Instruct 90B (Vision) when you need lower cost. Choose GPT-5.3 Codex (xhigh) when you need higher intelligence (54.0), faster output (72 tok/s).
Llama 3.2 Instruct 90B (Vision) is cheaper at $0.72/1M input tokens vs $1.75/1M for GPT-5.3 Codex (xhigh).
GPT-5.3 Codex (xhigh) is faster, producing output at 72 tok/s compared to Llama 3.2 Instruct 90B (Vision)'s 42 tok/s.
No, Llama 3.2 Instruct 90B (Vision) does not support image input. Neither model supports image input.
It depends on your priorities. GPT-5.3 Codex (xhigh) scores higher on intelligence (54.0), but Llama 3.2 Instruct 90B (Vision) may be better for specific use cases like budget-conscious projects or speed-critical applications.
Data last synced: March 26, 2026
| N/A |
| Max Output Tokens | N/A | N/A |
| Input Modalities | Text | Text |
| Output Modalities | Text | Text |
| Free Tier | No | No |