LLM Pricing

Cost guide

Output vs Input LLM API Pricing — Why Generation Costs 3–4× More

Published 2026-07-10 · Estimated read: 5 min

Output tokens cost 3.7× more than input tokens on average across 2,092 chat API models. That means your bill is dominated by generation cost even when your input volume looks larger. This guide breaks down the multiplier by provider, model type, and real workload examples.

1. The output/input pricing multiplier

Every LLM provider charges different rates for input and output tokens. While many developers focus on the headline input price, output tokens are typically 2–5× more expensive — and for reasoning models, the gap can reach 10× or more.

  • All chat models: 3.65× average output/input ratio
  • Non-reasoning models: 3.12× average
  • Reasoning models: 4.81× average
  • Total models analyzed: 2,092

The implication is straightforward: output cost drives the bill. Even when input volume is 2–3× larger, the higher output rate means generation accounts for most of the total. A model with cheap input but expensive output can cost more overall than a balanced alternative.

2. Per-provider breakdown

The output/input ratio varies dramatically by provider. Open-weight inference platforms like Fireworks AI typically offer near-1:1 ratios, while frontier providers like Anthropic and OpenAI consistently charge 4–5× for output.

Provider Avg output/input ratio Models
anyscale 1.00× 12
fireworks_ai 1.41× 243
nscale 1.62× 14
ovhcloud 1.72× 15
hyperbolic 1.84× 16
watsonx 1.85× 28
lambda_ai 2.01× 20
DeepSeek 2.12× 12
llamagate 2.19× 14
Groq 2.61× 11
oci 2.77× 29
Mistral 2.84× 51
wandb 3.00× 16
nebius 3.08× 27
gradient_ai 3.20× 10
vertex_ai-mistral_models 3.21× 19
Bedrock 3.27× 187
novita 3.28× 80
Scaleway 3.29× 14
deepinfra 3.30× 67
perplexity 3.50× 16
vercel_ai_gateway 3.54× 91
Fireworks AI 3.56× 33
baseten 3.59× 11
zai 3.60× 10
azure_ai 3.76× 65
Together AI 3.92× 32
xai 4.13× 38
OpenRouter 4.38× 93
Cloudflare 4.44× 22
Tensormesh 4.53× 10
Libertai 4.61× 11
bedrock_converse 4.62× 121
OpenAI 4.63× 103
replicate 4.63× 40
databricks 4.84× 25
moonshot 4.85× 22
Azure 4.91× 132
Snowflake 4.91× 11
gmi 4.93× 17
Bedrock Converse 5.00× 16
Anthropic 5.00× 23
vertex_ai-anthropic_models 5.00× 29
dashscope 5.41× 17
vertex_ai-language-models 5.95× 22
Google 6.38× 37
sambanova 13.69× 17

Fireworks AI leads with the lowest average ratio at 1.00×, making it a strong candidate for output-heavy workloads. At the other end, providers like Anthropic and OpenAI charge output at 4–5× input, reflecting the higher compute cost of generation.

3. Reasoning vs non-reasoning models

Reasoning (chain-of-thought) models amplify the output multiplier. These models generate internal reasoning tokens before producing the final answer, effectively increasing the output-to-input cost ratio.

  • Non-reasoning models: 3.12× average ratio (1,435 models)
  • Reasoning models: 4.81× average ratio (657 models)
  • Additional hidden tokens: Reasoning models often consume 30–40% more tokens internally through thinking chains, making the effective cost even higher than the ratio suggests

Our reasoning LLM pricing guide covers the full breakdown of thinking model costs.

4. Models with the cheapest output ratios

A few models buck the trend, charging near-1:1 or even cheaper output than input. These are worth considering for output-heavy workloads:

Model Provider Ratio Input / 1M Output / 1M
llama-guard-3-8b Cloudflare 0.06× $0.48 $0.03
Llama-4-Maverick-17B-128E-Instruct-FP8 azure_ai 0.25× $1.41 $0.35
Meta-Llama-3-70B-Instruct azure_ai 0.34× $1.10 $0.37
ai21.j2-mid-v1 Bedrock 1.00× $12.50 $12.50
ai21.j2-ultra-v1 Bedrock 1.00× $18.80 $18.80

At the other end of the spectrum, models with the highest output ratios (often specialized audio or multimodal routes) can reach 20–200×. Always check the ratio before committing to a model for generation-heavy workloads.

5. Real workload cost comparison

To see the multiplier in action, consider a standard production workload: 2,000 input tokens + 1,000 output token per request, 100,000 requests/month.

Scenario Input cost Output cost Total monthly Output share
Budget (2x ratio) $15.00 $15.00 $30.00 50%
Average (3.64x ratio) $100.00 $182.00 $282.00 65%
Premium (5x ratio) $600.00 $1500.00 $2100.00 71%

Even at the average 3.64× ratio, output cost accounts for 65% of the total bill. At the budget 2× ratio, it's still 50%. For premium 5× models, output dominates at 71%.

Use the calculator with your own token counts and the compare tool to see the real ratio for your shortlisted models.

6. Summary

  • Output tokens cost 3.7× more than input tokens on average
  • Non-reasoning models: 3.1×; reasoning models: 4.8×
  • Open-weight providers like Fireworks offer near-1:1 ratios; frontier providers charge 4–5×
  • Even with 2:1 input-to-output token volume, output drives 50–70%+ of the total cost

What to do next: Open the compare tool to check the output/input ratio for your shortlisted models, or use the calculator to estimate your workload's true output cost. The lowest-ratio models compare preset is a good starting point for output-heavy applications.