Match the model to the job, not to the benchmark.

Benchmarks rank models on tasks nobody is running in production. This tool matches a concrete job — real-time voice, long-document retrieval, structured extraction, high-volume classification — to the model characteristics that actually govern it, and shows the cost consequence of each choice.

The dimensions that decide it

• Time to first token — the only latency number that matters in voice.
• Instruction adherence under a long system prompt.
• Context window, and how quality degrades near the top of it.
• Refusal behaviour — over-refusal breaks agents in regulated trades.
• Cost per thousand tokens at your realistic conversation length.

Where benchmarks mislead most

Reasoning benchmarks reward models that think longer, which is precisely the wrong trait for a voice agent where a second of silence loses the caller. A model that scores worse on paper frequently produces a better conversation, and no public leaderboard will tell you that.

The routing pattern worth copying

Use a fast model for the conversational turn and a stronger one for the occasional hard call — a complex booking conflict or an ambiguous intent. Most production agents do not need one model; they need a cheap default and an expensive escape hatch that fires maybe five percent of the time.