Two Metrics, Two Different Questions
When developers talk about 'what's the most cost-efficient LLM', they often mean two completely different things. Tokens Per Dollar (TPD) answers 'how much compute do I get for $1' — useful when quality is held constant or doesn't matter. The Value column on TokenRate answers 'how much quality-adjusted compute do I get for $1' — useful when quality and price both matter. These are not interchangeable. A model with 10 million tokens per dollar at quality 15 has the same TPD ranking as a model with 10M tokens per dollar at quality 70 — but the second one is the only one you'd actually ship. This post walks through when to use each metric and how they show up on the calculator and Compare Prices tool.
Tokens Per Dollar: The 2024 Metric
Tokens Per Dollar is the legacy LLM cost metric. Math: 1,000,000 ÷ input cost per 1M tokens. A model at $0.10/1M input gives you 10M tokens per dollar; at $1/1M it gives 1M tokens per dollar; at $15/1M (Claude Opus 4) it gives 67,000. The metric was great in 2023–2024 when 'is this model production-grade' was usually a yes/no and you mostly cared about throughput. In 2026 it's incomplete: every provider now ships models from $0.05/1M input up to $20/1M input, and quality varies enormously across that range. Sorting by TPD alone surfaces budget-tier models that may or may not be production-ready. For TPD-specific context, see tokens per dollar comparison 2026.
Value Column: The 2026 Metric
TokenRate's Value column is quality score ÷ input cost per 1M tokens. Higher is better. A model at quality 73 / $0.55 input = value 133. At quality 80 / $15 input = value 5.3. The Value metric is the right default for production routing decisions because it penalizes both 'cheap-and-junk' (low quality) and 'expensive-and-incremental' (high price). When you sort the calculator by 'best value', you see the highest-quality model that fits any given price point — which is exactly the production question. The formula is also workload-portable: a multi-step agent on a $200/mo budget gets the same Value ranking as a single-shot classifier on a $20,000/mo budget. For broader context, see quality per dollar LLM ranking 2026 and why the cheapest LLM isn't always the best value.
When TPD Still Wins
Tokens Per Dollar is the right metric in three cases. (1) Quality is genuinely fungible — synthetic data generation (sampling errors average out), embedding-adjacent classification (small fixed output set), pre-filter triage stages. (2) You've already constrained to a single quality tier and want to maximize throughput within it — e.g., 'cheapest credible Top (75+) model' or 'cheapest credible fast-tier model'. (3) You're benchmarking the absolute cost floor — research, budgeting, what's-possible-at-the-bottom. For these cases, sort the calculator by 'cheapest' (which reads as 'highest TPD') and the result is what you want. For everything else, use the Value column.
Side-by-Side: A Worked Example
Take three models. Llama 3.2 1B: quality 15, $0.05 input → TPD 20,000,000, Value 300. Gemini 2.5 Flash-Lite: quality 55, $0.075 input → TPD 13,300,000, Value 733. Claude Sonnet 4.7: quality 80, $3 input → TPD 333,000, Value 27. By TPD: Llama wins by 50%. By Value: Gemini Flash-Lite wins by 2.4x over Llama and by 27x over Sonnet — the production-best pick across a wide range of workloads. By raw quality: Sonnet wins. By value-at-flagship-tier (filter to Quality 75+): Sonnet is the only candidate. Each metric answers a different question. The Filter panel plus the right sort gives you the answer to whichever question your workload actually needs. See how to pick an LLM by quality score and cost for the full workflow.