BridgeBenchBridgeBench
Live

Speed

How fast can each AI coding model generate code? Throughput and TTFT measured separately for accurate, real-world numbers that matter for vibe coding and agentic coding workflows.

Updated 2026-03-27

Throughput

Sustained tok/s measured first-to-last token on 1,500+ token outputs.

TTFT

Time to first token on short prompts. How fast the model starts responding.

Methodology

Median across 3+ runs per prompt. Short and long outputs measured separately.

RankModelThroughputTTFT
Grok 4.20 (Non-Reasoning)

x-ai/grok-4.20

243.3t/s
1999ms
Grok 4.20 Reasoning

x-ai/grok-4.20-reasoning

237.7t/s
1497ms
Gemini 3.1 Pro

google/gemini-3.1-pro-preview

122.2t/s
7608ms
4Claude Sonnet 4.6

anthropic/claude-sonnet-4-6

95.3t/s
1207ms
5Claude Opus 4.6

anthropic/claude-opus-4-6

92.2t/s
1922ms
6GPT-5.4

openai/gpt-5.4

88t/s
397ms
7GLM-5

openrouter/z-ai/glm-5

73.3t/s
962ms
8MiniMax M2.7

minimax/MiniMax-M2.7

68.2t/s
6150ms
9Grok 4

x-ai/grok-4

61.4t/s
3684ms
10MiMo-V2-Pro

openrouter/xiaomi/mimo-v2-pro

57.5t/s
7791ms
11GLM 5.1

z-ai/glm-5.1

44.3t/s
2353ms

11 modelstested · Ranked by median throughput · More models coming soon