Inference Index
Compare

DeepSeek R2 vs Llama 4 Behemoth

DeepSeek vs Meta. Specs, benchmarks, and real per-task cost — all in one page.

Verdict

DeepSeek R2 is ~4.0× cheaper on a 3:1 input:output blend.

DeepSeek
DeepSeek R2

DeepSeek’s reasoning variant. Competes with o4 on math at a fraction of the cost.

Frontier math at open-weights pricingLong CoT traces
Meta
Llama 4 Behemoth

Meta’s open-weights flagship. 405B params, fully open license, runs on every major inference provider.

Best open-weights modelRuns anywhereTransparent training
Pricing
Input / 1M$0.55$2.50
Output / 1M$2.19$8.00
Context128K256K
Max output64K16K
LicenseOpen weightsOpen weights
Released2026-01-152025-12-10
Benchmarks
LMSYS ELO1342.0
MMLU Pro87.887.1
HumanEval90.189.3
SWE-Bench52.4
MATH94.181.5
GPQA71.256.2
IFEval88.0
Per-task cost
Summarize a 1-hour meeting transcript$0.0093$0.041
Review a 500-line pull request$0.0088$0.036
Answer a customer support ticket$0.0034$0.014
Extract structured data from a resume$0.0049$0.021
Debug a stack trace with context$0.0077$0.031

Per-call cost using published token counts for each task. Real-world prompts vary.