Output tokens are most of your bill, and the ratio varies by 12×
Across 209 billable models the median output rate is 4.0× the input rate (mean 4.5×). The steepest is Qwen3 30B A3b Think at 12×; 16 models charge the same for both. That spread is the single biggest reason two teams running the same model can report wildly different unit costs — and it is invisible on any price table that only prints the input column.
https://aicomp.ai/v1).
Create one free →
curl https://aicomp.ai/v1/models \
-H "Authorization: Bearer sk-your-gateway-key"A JSON list of model IDs means the key is good. Invalid token means it was copied wrong.
The distribution
| Percentile | Output ÷ input | Output share at 1:1 volume |
|---|---|---|
| cheapest 10% | 2.00× | 67% |
| 20% | 3.00× | 75% |
| 30% | 3.14× | 76% |
| 40% | 4.00× | 80% |
| median | 4.00× | 80% |
| 60% | 5.00× | 83% |
| 70% | 5.00× | 83% |
| 80% | 6.00× | 86% |
| 90% | 8.00× | 89% |
Read the last column as: if you send equal numbers of input and output tokens, this is the share of the bill that goes to output. At the median ratio that is already 80%.
The two ends of the range
Steepest — output punishes you
| Model | Vendor | Input | Output | Ratio | Output share (1:3) |
|---|---|---|---|---|---|
| Qwen3 30B A3b Think | Alibaba | $0.1 | $1.2 | 12.00× | 97% |
| Qwen3 30B A3b Thinking 2507 | Alibaba | $0.1 | $1.2 | 12.00× | 97% |
| Doubao Seed 1 6 251015 Search | ByteDance | $0.4 | $4 | 10.00× | 97% |
| Seed Oss 36B Instruct | ByteDance | $0.6 | $6 | 10.00× | 97% |
| Qwen3 235B A22b Thinking 2507 | Alibaba | $0.115 | $1.15 | 10.00× | 97% |
| Gemini 2.5 Flash | $0.15 | $1.251 | 8.34× | 96% | |
| Gemini 3.5 Flash Lite | $0.15 | $1.25 | 8.33× | 96% | |
| Qwen Flash | Alibaba | $0.025 | $0.2 | 8.00× | 96% |
Flattest — the ratio barely exists
| Model | Vendor | Input | Output | Ratio | Output share (1:3) |
|---|---|---|---|---|---|
| Qwen 72B | Alibaba | $1 | $1 | 1.00× | 75% |
| Doubao Lite 32K | ByteDance | $1 | $1 | 1.00× | 75% |
| Doubao Lite 4K | ByteDance | $1 | $1 | 1.00× | 75% |
| Doubao Lite 128K | ByteDance | $2 | $2 | 1.00× | 75% |
| Doubao Pro 32K | ByteDance | $2 | $2 | 1.00× | 75% |
| Doubao Pro 4K | ByteDance | $2 | $2 | 1.00× | 75% |
| Doubao Pro 128K | ByteDance | $12 | $12 | 1.00× | 75% |
| DeepSeek OCR | DeepSeek | $0.108 | $0.108 | 1.00× | 75% |
What the ratio does to a real mix
Now the same arithmetic at the steep end. Qwen3 30B A3b Think bills $0.1 per 1M in and $1.2 per 1M out — a 12× ratio. On that same 1:3 mix, output carries 97% of the cost. Halve your prompt length and you have barely touched the invoice.
This is the practical consequence: optimisation effort follows the money, and the money is on the output side. Shorter system prompts, tighter retrieval, aggressive deduplication — all worth doing, all smaller than people assume once output dominates.
When flat pricing is worth paying for
16 of these 209 models bill input and output at an identical rate. On them the mix stops mattering: whatever your ratio turns out to be, the blended rate equals the single listed rate. That is worth real money when:
- You cannot predict the mix. User-generated prompts, extraction over documents of varying length, anything with retries.
- You are budgeting before launch. A flat rate turns a two-variable estimate into one.
- Your output is long and your input is short. Flat pricing is effectively a discount on the expensive side — check the blended figure before dismissing a model whose input rate looks high.
The trade-off is honest: on read-heavy work you give up the cheap-input advantage. Flat is predictability, not a bargain.
Three things that actually cut output cost
- Cap generation length. Always set
max_tokens. An unbounded generation is an unbounded bill, and one stuck loop can outweigh a month of prompt tuning. - Ask for structure, not prose. A JSON schema with fixed fields costs a predictable number of tokens; "explain your reasoning" does not.
- Split the work by ratio. Reasoning on a strong model, drafting on a cheap one. The expensive model's output is short, which is exactly where its rate hurts least.
FAQ
Why is output more expensive than input?
Generation costs compute per token produced; reading the prompt is a parallel, one-pass operation. Vendors price them separately and the ratio has settled in a wide band — a median of 4.0x across 209 models here, with a long tail up to 12x. It is not a penalty, it is the actual cost shape of autoregressive decoding.
What is a normal output-to-input ratio?
Median 4.0x, mean 4.5x across 209 models. Roughly 32 models sit at or below 2x and 31 sit at 8x or above, so the band is genuinely wide — which is exactly why you should check the ratio for the models you shortlist rather than assume an industry average.
Are there models that charge the same for input and output?
Yes — 16 of the 209 models here bill both at an identical rate. For a workload whose token mix you cannot predict (user-generated prompts, variable-length extraction, anything with a retry loop), flat pricing removes the ratio from your forecast entirely. You lose the cheap-input advantage on read-heavy work and gain predictability on everything else.
How much of my bill is output tokens?
On a 1:3 agent mix and a median 4.0x ratio, output carries about 92% of the cost while being 75% of the tokens. At the steep end (Qwen3 30B A3b Think, 12x) it is closer to 97%. This is why cutting input cost — shorter prompts, tighter retrieval — has a much smaller effect than people expect once output dominates.
Does prompt caching change which side matters?
It makes the output side matter more. A cache hit bills at roughly 0.1x base input on Anthropic, OpenAI's GPT-5.6 generation and Gemini 2.5+, which shrinks the input line and leaves output as an even larger share of the total. Optimise input if your prompts are long and repetitive; optimise output if they are not.
How do I reduce output cost specifically?
Cap max_tokens so a runaway generation cannot bill unbounded; ask for structured, bounded output (JSON with a schema rather than prose); and route long drafts to a cheaper model while keeping the expensive one for the reasoning step. Streaming does not reduce cost — it only changes when you see the tokens.
Are these official vendor rates?
They are gateway rates we bill, USD per 1M tokens, captured 2026-09-20. 13 of the 209 models carry a published vendor list rate we can cross-check and are marked verified; the rest are listed at face value. Ratios are computed from those two rates and are not an extra charge.