Compare
GPT-5 Mini vs Llama 4 Behemoth
OpenAI vs Meta. Specs, benchmarks, and real per-task cost — all in one page.
Verdict
Llama 4 Behemoth leads on LMSYS ELO (1342 vs 1288). GPT-5 Mini is ~4.4× cheaper on a 3:1 input:output blend. Llama 4 Behemoth is open-weights, the other is proprietary.
OpenAI
GPT-5 MiniGPT-5’s smaller, cheaper sibling. Optimized for chat and lightweight agent tasks.
Very low latencyCheap enough for mass consumer apps
Meta
Llama 4 BehemothMeta’s open-weights flagship. 405B params, fully open license, runs on every major inference provider.
Best open-weights modelRuns anywhereTransparent training
Pricing
Input / 1M$0.50$2.50
Output / 1M$2.00$8.00
Context256K256K
Max output16K16K
LicenseProprietaryOpen weights
Released2026-01-212025-12-10
Benchmarks
LMSYS ELO1288.01342.0
MMLU Pro78.587.1
HumanEval85.589.3
SWE-Bench—52.4
MATH79.881.5
GPQA—56.2
IFEval85.488.0
Per-task cost
Summarize a 1-hour meeting transcript$0.0085$0.041
Review a 500-line pull request$0.0080$0.036
Answer a customer support ticket$0.0031$0.014
Extract structured data from a resume$0.0045$0.021
Debug a stack trace with context$0.0070$0.031
Per-call cost using published token counts for each task. Real-world prompts vary.