Model verbosity: Save 92–95% on LLM API Costs
Does model verbosity save money on LLM API costs?
Yes, for the right workload shape: 92–95% based on Measured verbosityIndex spans 0.63×–12.95× across 24 graded models. Swapping the most verbose model for a median one (index 1.00×) saves the low end; swapping it for the least verbose covered model saves the high end. It does not apply universally — see the applicability and tradeoff sections below before using it.
Batch 40 · server-rendered decision evidence · verified 2026-08-27
Model-verbosity and accepted-answer cost evidence
Frozen inputs, formulas, provenance, and decision boundaries are visible in the initial HTML. Unsupported evidence fails closed as Unavailable.
Task-stratified output distribution
Formula / scoring rule: Output efficiency is task-local: compare P50/P90 tokens, grader pass, retries, and output bill on matched runs only.
Provenance: Frozen coding, writing, extraction, and reasoning fixtures; model/run IDs and grader results retained.
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | State |
|---|---|---|---|---|
codingbatch40-model-verbosity-m1-r1 | run C-40; P50 620 / P90 1,140 output tokens; grader 18/20 | Output bill under $15/M = $0.0093 at P50. | Do not transfer coding verbosity to writing. | MATCHED — task-local evidence. |
writingbatch40-model-verbosity-m1-r2 | run W-40; P50 880 / P90 1,900; grader Unavailable — matched quality run is absent | Unavailable — quality denominator is absent | Token distribution alone cannot select a model. | Unavailable — quality denominator is absent |
reasoningbatch40-model-verbosity-m1-r3 | run R-40; P50 1,420 / P90 3,400; hidden reasoning usage | Visible output is measured; hidden usage Unavailable — not returned | No total-bill claim from visible output only. | Unavailable — not returned |
Module citation: All AI Ask verbosity dataset.
Intervention canary and truncation surface
Formula / scoring rule: Net tokens = emitted + continuation + retry; intervention passes only if grader floor holds and finish state is not length.
Provenance: Identical prompts under unchanged, concise, schema, low/medium/high controls and hard caps.
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | State |
|---|---|---|---|---|
concise directivebatch40-model-verbosity-m2-r1 | coding fixture; baseline 620 output; canary 410; finish=stop | Token reduction 210; grader delta Unavailable — matched canary grader is absent | Shorter output is not accepted without grader parity. | Unavailable — matched canary grader is absent |
structured schemabatch40-model-verbosity-m2-r2 | extraction; 510 output; schema accepted; finish=stop | Fields 40/40; output reduced to 390; retry 0. | Schema acceptance is a hard gate before cost comparison. | PASS — fixture-specific. |
hard capbatch40-model-verbosity-m2-r3 | writing cap 512; emitted 512; finish=length; continuation 1 | Net output 1,040; apparent reduction reverses after continuation. | A length finish is not a successful concise answer. | TRUNCATED — excluded from savings claim. |
Module citation: OpenAI model controls.
Cost-per-accepted-answer frontier
Formula / scoring rule: Cost/accepted answer = (input + output + continuation + retry charges) / accepted answers; quality floor is user-supplied.
Provenance: Output rates sourced separately; acceptance floor and review cost remain explicit user variables.
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | State |
|---|---|---|---|---|
terse candidatebatch40-model-verbosity-m3-r1 | P50 410 output; rate $15/M; acceptance floor 90% | Output subtotal $0.00615/answer; acceptance Unavailable — not measured on matched run | Cannot enter frontier without acceptance denominator. | Unavailable — not measured on matched run |
verbose candidatebatch40-model-verbosity-m3-r2 | P50 880 output; same rate; acceptance 94% scenario | Output subtotal $0.0132; cost per accepted Unavailable — scenario is not observed | Scenario acceptance cannot be presented as quality evidence. | Unavailable — scenario is not observed |
retry crossoverbatch40-model-verbosity-m3-r3 | terse retry share 12%; verbose retry share 4%; user floor 92% | Crossover requires measured pass rates and retry bills: Unavailable — inputs are incomplete | Do not call terse cheaper when retries erase token savings. | Unavailable — inputs are incomplete |
Module citation: All AI Ask pricing registry.
How it works
Output tokens, not input tokens, are the lever here. We run every priced model against the same graded prompts and measure how many output tokens it actually spends per answer; the ratio to the group median across models that ran the same test is its verbosityIndex. A model at 1.6× talks 60% more than the median model to answer the identical question — and every one of those extra tokens is billed at the output rate, which is usually 3-5× the input rate. Swapping a high-verbosity model for a lower one on the same job cuts the output-cost side of the bill directly, without changing volume, caching, or batching.
Where it applies
Any output-token-heavy workload. Does nothing for jobs bottlenecked on input tokens (large retrieved context, long documents) since verbosity only scales the output side.
Verbosity range by provider
| Provider | Measured models | Verbosity range |
|---|---|---|
| OpenAI | 2 | 0.78×–1.04× |
| Anthropic | 1 | 0.96×–0.96× |
| 2 | 0.63×–0.87× | |
| xAI | 1 | 1.41×–1.41× |
| DeepSeek | 2 | 2.59×–3.30× |
| Mistral | 4 | 0.79×–0.93× |
| Groq | 2 | 1.82×–2.93× |
| Cerebras | 2 | 2.32×–7.53× |
| Amazon | 2 | 0.76×–0.91× |
| Z.ai | 1 | 3.87×–3.87× |
| Meta | 1 | 12.95×–12.95× |
The tradeoff
A less verbose model is not free to swap to — it is a different model, with a different accuracy profile on your specific task. The verbosity index measures how many tokens a model spends per answer, not whether that answer is correct; a terser model that gets more graded runs wrong costs you in rework even if its raw token bill is lower. Verify accuracy on your own task before swapping, using the graded results on /best-llm-for, not the cost table alone. Verbosity is also measured per task family — a model terse on summarization can still be verbose on code generation, so the index for one job does not transfer to another.
Worked example
For llm chatbot (growing conversation history in, short reply out) at 200,000 calls/month, GPT-5.4 Mini runs $675/month at list price and $675/month once verbosity is priced in. See the full ranked table for this workload →