DeepSeek V4 Pro 1.6T
8K/1K
Hyperscaler cost ↓ Lower is better Source: InferenceX & SemiAnalysis Market July 2026 AI Cloud TCO Model
| Model · Scenario | B200 · Reference | MI355X | B300 | GB200 NVL72 | GB300 NVL72 |
|---|---|---|---|---|---|
DeepSeek V4 Pro 1.6T8K/1K | vLLM · FP4 | $0.8198% more expensive than B200 SGLang · FP4 | $0.31958% cheaper than B200 Dynamo SGLang · FP4 | $2.668252% more expensive than B200 Dynamo vLLM · FP4 | $0.18076% cheaper than B200 Dynamo SGLang · FP4 |
DeepSeek V4 Pro 1.6TLong Context Multi-Turn Realistic Agentic Scenario (AgentX) | vLLM · FP4 | — cannot reach @100 | $0.07346% cheaper than B200 vLLM · FP4 | — no data for this scenario | $0.03574% cheaper than B200 Dynamo vLLM · FP4 |
Kimi K3 2.8TLong Context Multi-Turn Realistic Agentic Scenario (AgentX) | — cannot reach @100 | — no data for this scenario | $0.197No B200 baseline to compare against vLLM · FP4 | — cannot reach @100 | — no data for this scenario |
MiniMax M3 428B8K/1K | vLLM · FP4 | $0.14596% more expensive than B200 vLLM · FP4 | $0.08819% more expensive than B200 vLLM · FP4 | $0.295300% more expensive than B200 Dynamo vLLM · FP8 · STP | $0.10441% more expensive than B200 Dynamo vLLM · FP8 |
MiniMax M3 428BLong Context Multi-Turn Realistic Agentic Scenario (AgentX) | vLLM · FP4 | — no data for this scenario | — no exact @100 result | — no data for this scenario | — no data for this scenario |
GLM5.2Long Context Multi-Turn Realistic Agentic Scenario (AgentX) | SGLang · FP4 | — no data for this scenario | $0.06118% cheaper than B200 SGLang · FP4 | — no data for this scenario | — no data for this scenario |
Qwen3.5 397B8K/1K | SGLang · FP4 | $0.09957% more expensive than B200 SGLang · FP4 | $0.061About the same cost as B200 SGLang · FP4 | $0.11277% more expensive than B200 Dynamo SGLang · FP8 | $0.133111% more expensive than B200 Dynamo SGLang · FP8 |
Qwen3.5 397BLong Context Multi-Turn Realistic Agentic Scenario (AgentX) | SGLang · FP4 | — no data for this scenario | $0.01145% more expensive than B200 SGLang · FP4 | — no data for this scenario | $0.00929% more expensive than B200 Dynamo SGLang · FP4 |
DeepSeek R1 0528 671BMaintenance8K/1K | SGLang · FP4 | $0.11536% cheaper than B200 MoRI SGLang · FP4 | $0.514187% more expensive than B200 SGLang · FP8 | $0.531196% more expensive than B200 Dynamo SGLang · FP4 · STP | $0.490174% more expensive than B200 Dynamo SGLang · FP4 · STP |
Kimi K2.5/2.6/2.7-Code 1TDeprecated8K/1K | vLLM · FP4 · STP | $0.2899% cheaper than B200 vLLM · FP4 · STP | $0.36916% more expensive than B200 vLLM · FP4 · STP | $0.38320% more expensive than B200 Dynamo vLLM · FP4 · STP | $0.52866% more expensive than B200 Dynamo vLLM · FP4 · STP |
GLM5/5.1 744BDeprecated8K/1K | SGLang · FP4 | — cannot reach @100 | $0.41231% more expensive than B200 SGLang · FP4 | $0.36616% more expensive than B200 Dynamo SGLang · FP4 | $0.50460% more expensive than B200 Dynamo SGLang · FP4 |
gpt-oss 120BDeprecated8K/1K | vLLM · FP4 · STP | $0.0188% more expensive than B200 vLLM · FP4 · STP | — no data for this scenario | — no data for this scenario | — no data for this scenario |
MiniMax M2.5/2.7 230BDeprecated8K/1K | vLLM · FP4 · STP | $0.162129% more expensive than B200 vLLM · FP4 · STP | $0.09230% more expensive than B200 vLLM · FP4 · STP | $0.08216% more expensive than B200 Dynamo vLLM · FP4 · STP | $0.09737% more expensive than B200 Dynamo vLLM · FP4 · STP |
Llama 3.3 70B InstructDeprecated8K/1K | vLLM · FP4 · STP | $0.674251% more expensive than B200 vLLM · FP4 · STP | — no data for this scenario | — no data for this scenario | — no data for this scenario |
8K/1K
Long Context Multi-Turn Realistic Agentic Scenario (AgentX)
Long Context Multi-Turn Realistic Agentic Scenario (AgentX)
8K/1K
Long Context Multi-Turn Realistic Agentic Scenario (AgentX)
Long Context Multi-Turn Realistic Agentic Scenario (AgentX)
8K/1K
Long Context Multi-Turn Realistic Agentic Scenario (AgentX)
8K/1K
8K/1K
8K/1K
8K/1K
8K/1K
8K/1K
— = no result. ∞ = B200 baseline unavailable.
If a chip does not have FP4 spec decoding available, the next best available configuration is used.