PRICE~moonshotai/kimi-latest cached_input_per_mtok decreased from 0.80 to 0.291m PRICE~moonshotai/kimi-latest output_per_mtok decreased from 13.00 to 11.361m PRICE~moonshotai/kimi-latest input_per_mtok decreased from 0.99 to 0.751m PRICE~deepseek/deepseek-pro-latest cached_input_per_mtok increased from 0.0042 to 0.101m PRICE~deepseek/deepseek-pro-latest output_per_mtok increased from 0.40 to 4.201m PRICE~deepseek/deepseek-pro-latest input_per_mtok decreased from 0.13 to 0.131m PRICE~moonshotai/kimi-latest cached_input_per_mtok increased from 0.29 to 0.8032m PRICE~moonshotai/kimi-latest output_per_mtok increased from 11.36 to 13.0032m PRICE~moonshotai/kimi-latest input_per_mtok increased from 0.75 to 0.9932m PERFORMANCEGLM-5.2 → Scaleway: TTFT ↑ 20%1h PERFORMANCEDeepSeek V4 Flash → Scaleway: TTFT ↑ 35%1h PERFORMANCEMistral Medium 3.5 → Scaleway: TTFT ↓ 21%1h PERFORMANCEDeepSeek V4 Flash → Scaleway: throughput ↓ 20%1h PERFORMANCEGPT-OSS 120B → Scaleway: throughput ↑ 31%1h PERFORMANCEGPT-OSS 120B → Scaleway: TTFT ↓ 24%1h PERFORMANCEGPT-OSS 120B → Together AI: throughput ↑ 20%1h PERFORMANCEKimi K3 via OpenRouter: TTFT ↓ 49%1h PERFORMANCEGLM-5.2 via OpenRouter: reliability recovered 96.8% → 100.0%1h PERFORMANCEKimi K3 via Cortecs: throughput ↑ 184%1h PERFORMANCEKimi K3 via Cortecs: TTFT ↓ 42%1h PERFORMANCEGLM-5.2 via OpenRouter: throughput ↓ 49%1h PERFORMANCEGLM-5.2 via OpenRouter: TTFT ↑ 121%1h PERFORMANCEMiniMax M3 via OpenRouter: reliability recovered 96.8% → 100.0%1h PERFORMANCEDeepSeek V4 Pro via Cortecs: throughput ↑ 27%1h PERFORMANCEGPT-OSS 120B via OpenRouter: throughput ↓ 22%1h PERFORMANCEGLM-4.7 via OpenRouter: TTFT ↓ 78%1h PERFORMANCEGPT-OSS 20B → Groq: TTFT ↑ 24%1h PERFORMANCEKimi K3 → Together AI: TTFT ↓ 34%1h PERFORMANCEDeepSeek V4 Flash via OpenRouter: throughput ↓ 17%1h PERFORMANCELlama 3.3 70B via Cortecs: TTFT ↓ 51%1h PERFORMANCEQwen3 235B via OpenRouter: throughput ↑ 37%1h PERFORMANCEQwen3 235B via OpenRouter: TTFT ↓ 29%1h PERFORMANCEDeepSeek V4 Flash → Together AI: throughput ↑ 19%1h PERFORMANCEGLM-5.3 via OpenRouter: throughput ↓ 17%1h PERFORMANCEGPT-OSS 20B via Cortecs: throughput ↑ 28%1h PERFORMANCELlama 3.3 70B → Together AI: TTFT ↓ 33%1h PRICE~moonshotai/kimi-latest cached_input_per_mtok decreased from 0.80 to 0.291h PRICE~moonshotai/kimi-latest output_per_mtok decreased from 13.00 to 11.361h PRICE~moonshotai/kimi-latest input_per_mtok decreased from 0.99 to 0.751h PRICE~deepseek/deepseek-v4-flash-latest cached_input_per_mtok decreased from 0.0077 to 0.00111h

PRICE~moonshotai/kimi-latest cached_input_per_mtok decreased from 0.80 to 0.291m PRICE~moonshotai/kimi-latest output_per_mtok decreased from 13.00 to 11.361m PRICE~moonshotai/kimi-latest input_per_mtok decreased from 0.99 to 0.751m PRICE~deepseek/deepseek-pro-latest cached_input_per_mtok increased from 0.0042 to 0.101m PRICE~deepseek/deepseek-pro-latest output_per_mtok increased from 0.40 to 4.201m PRICE~deepseek/deepseek-pro-latest input_per_mtok decreased from 0.13 to 0.131m PRICE~moonshotai/kimi-latest cached_input_per_mtok increased from 0.29 to 0.8032m PRICE~moonshotai/kimi-latest output_per_mtok increased from 11.36 to 13.0032m PRICE~moonshotai/kimi-latest input_per_mtok increased from 0.75 to 0.9932m PERFORMANCEGLM-5.2 → Scaleway: TTFT ↑ 20%1h PERFORMANCEDeepSeek V4 Flash → Scaleway: TTFT ↑ 35%1h PERFORMANCEMistral Medium 3.5 → Scaleway: TTFT ↓ 21%1h PERFORMANCEDeepSeek V4 Flash → Scaleway: throughput ↓ 20%1h PERFORMANCEGPT-OSS 120B → Scaleway: throughput ↑ 31%1h PERFORMANCEGPT-OSS 120B → Scaleway: TTFT ↓ 24%1h PERFORMANCEGPT-OSS 120B → Together AI: throughput ↑ 20%1h PERFORMANCEKimi K3 via OpenRouter: TTFT ↓ 49%1h PERFORMANCEGLM-5.2 via OpenRouter: reliability recovered 96.8% → 100.0%1h PERFORMANCEKimi K3 via Cortecs: throughput ↑ 184%1h PERFORMANCEKimi K3 via Cortecs: TTFT ↓ 42%1h PERFORMANCEGLM-5.2 via OpenRouter: throughput ↓ 49%1h PERFORMANCEGLM-5.2 via OpenRouter: TTFT ↑ 121%1h PERFORMANCEMiniMax M3 via OpenRouter: reliability recovered 96.8% → 100.0%1h PERFORMANCEDeepSeek V4 Pro via Cortecs: throughput ↑ 27%1h PERFORMANCEGPT-OSS 120B via OpenRouter: throughput ↓ 22%1h PERFORMANCEGLM-4.7 via OpenRouter: TTFT ↓ 78%1h PERFORMANCEGPT-OSS 20B → Groq: TTFT ↑ 24%1h PERFORMANCEKimi K3 → Together AI: TTFT ↓ 34%1h PERFORMANCEDeepSeek V4 Flash via OpenRouter: throughput ↓ 17%1h PERFORMANCELlama 3.3 70B via Cortecs: TTFT ↓ 51%1h PERFORMANCEQwen3 235B via OpenRouter: throughput ↑ 37%1h PERFORMANCEQwen3 235B via OpenRouter: TTFT ↓ 29%1h PERFORMANCEDeepSeek V4 Flash → Together AI: throughput ↑ 19%1h PERFORMANCEGLM-5.3 via OpenRouter: throughput ↓ 17%1h PERFORMANCEGPT-OSS 20B via Cortecs: throughput ↑ 28%1h PERFORMANCELlama 3.3 70B → Together AI: TTFT ↓ 33%1h PRICE~moonshotai/kimi-latest cached_input_per_mtok decreased from 0.80 to 0.291h PRICE~moonshotai/kimi-latest output_per_mtok decreased from 13.00 to 11.361h PRICE~moonshotai/kimi-latest input_per_mtok decreased from 0.99 to 0.751h PRICE~deepseek/deepseek-v4-flash-latest cached_input_per_mtok decreased from 0.0077 to 0.00111h

Research

Independent analysis of AI inference performance, pricing and infrastructure, based on InferenceBench data.

ENGINEERING STUDY

The P95 Problem

Measured p95/p50 TTFT ratios span 1.5× to 7.2× across providers.

9k runs · Aug 22, 2026 ENGINEERING STUDY

Tool Calling Is an Infrastructure Property

On DeepSeek V4 Flash, measured tool-call reliability spans 0.0%–100.0% across 5 providers serving identical weights.

9k runs · Aug 22, 2026 ENGINEERING STUDY

Structured Output: Trust, but Validate

Measured structured-output validity spans 91.7%–100.0% across providers.

9k runs · Aug 22, 2026

Support independent AI infrastructure research.

Benchmark Partners · Become a Founding Partner →

AI inference research & market analysis · InferenceBench