NEW MODELQwen3-235B now available on Nebius34mPRICENextbit output price ↓ 12%51mNEW REGIONFireworks adds EU inference endpoint1hBENCHMARKDeepSeek-V3 re-benchmarked across 8 providers2hPERFORMANCECerebras throughput ↑ 12% this week2hNEW PROVIDERTensorix added to the registry3hINCIDENTGroq reliability dipped below 98.6%4hMODELLlama 3.3 70B live on Scaleway EU4hCAPABILITYMistral enables prompt caching5hPERFORMANCENebius overall score ↑ 1.6 in 30 days6hREGIONOVHcloud expands to second EU region7hBENCHMARK12,400 new benchmark runs on Qwen3-235B7hPRICEDeepInfra input price ↓ 8%8hCAPABILITYTogether batch API benchmarked8h

Benchmarks by use case

Different workloads weight the benchmark differently — the weighting is always visible, and every use case has its own deep benchmark.

AgentsBenchmarked for tool reliability, structured output, multi-step execution, latency and production reliability.Tool reliability · Structured output · Multi-step execution · Production reliabilityCodingBenchmarked for tool-driven editing loops, structured patches, long-context handling and latency.Tool reliability · Structured edits · Long context · LatencyRealtimeBenchmarked for time-to-first-token, p95 tail latency, streaming stability and availability.TTFT · p95 latency · Streaming stability · AvailabilityVoiceBenchmarked for first-token latency, inter-token pacing, stream integrity and availability under load.TTFT · Inter-token latency · Stream integrity · AvailabilityRAGBenchmarked for structured grounding, context handling, caching behaviour and cost at retrieval scale.Structured output · Context handling · Prompt caching · CostStructured outputBenchmarked for JSON validity, schema adherence, enum and nested-object correctness and truncation behaviour.Schema adherence · JSON validity · Reliability · LatencyBatchBenchmarked for throughput, cost per token, batch API support and sustained reliability.Throughput · Cost · Batch API · ReliabilityLong contextBenchmarked for context handling, caching, throughput on large prompts and price at high input volume.Context handling · Prompt caching · Input price · Throughput