Loading benchmark data…
EyesInAI·Loading live benchmark data
Live benchmark results across all tested AI models. Sortable by cost, latency, throughput.
microsoft/phi-4 dominates cost-quality frontier at $0.022/1k; retire sambanova and fireworks models.
Cerebras zai-glm-4.7 passed 5 of 7 tests before hitting rate limits; groq/openai-gpt-oss-120b achieved 57% pass rate across 7 tests; sambanova/MiniMax-M2.7 failed entirely with 20s+ timeouts. Pareto frontier dominated by openrouter/microsoft/phi-4 (100% pass, $0.022/1k tokens). Fireworks kimi-k2p5 model not found (504 retired). Most surprising: bytedance-seed/seed-2.0-mini achieves 317ms median latency—fastest across all tests.
openrouter/microsoft/phi-4 is the sole Pareto-optimal model: 100% pass rate, $0.022/1k tokens, 406ms latency. google/gemma-3-12b-it ($0.0237/1k, 100% pass) trades negligible cost for 2% premium but offers 1388.9 tps throughput. No other model achieves comparable cost-quality balance.
For tool_use: route to together/openai-gpt-oss-120b (100% pass, cheapest). For latency-sensitive tasks (ping, reasoning, code): use openrouter/microsoft/phi-4 (352–385ms). For throughput (speed test): use openrouter/google/gemma-3-12b-it (1388.9 tps). For cost optimization on context retrieval: phi-4 at $0.0183/1k cost-per-call.
bytedance-seed/seed-2.0-mini achieves 317ms median latency—lowest across all 100+ models tested. openrouter/microsoft/phi-4 and google/gemma-3-12b-it both deliver 100% pass rates at extreme cost efficiency ($0.022–0.024/1k tokens).
$0.022/1k runs at 100% pass rate
$0.0237/1k runs at 100% pass rate
lowest median latency across all tests at 317ms
Migrate all workloads to openrouter/microsoft/phi-4 (100% pass rate, $0.022/1k tokens, 406ms avg latency). For latency-critical paths, use google/gemma-3-12b-it ($0.0237/1k, 1388.9 tps). Retire sambanova MiniMax-M2.7 and fireworks kimi-k2p5 immediately.