Summaries are mostly input
A summary reads a lot and writes a little, so the input price matters more here than on most jobs. The example is 10,000 documents a month, about 3,000 tokens in and 300 tokens out each.
10,000 summaries on each model
List price, no caching, no batch.
| Model | Input | Output | Per month |
|---|---|---|---|
| Claude Sonnet 5.5 | $60.00 | $30.00 | $90.00 |
| Claude Haiku 5.5 | $3.00 | $1.50 | $4.50 |
| GPT-5.5 | $150 | $90.00 | $240 |
| GPT-6.1 Terra | $30.00 | $15.00 | $45.00 |
| GPT-6 Luna | $3.00 | $1.50 | $4.50 |
| GPT-5 mini | $7.50 | $6.00 | $13.50 |
| GPT-5 nano | $1.50 | $1.20 | $2.70 |
| Gemini 3.8 Flash | $22.50 | $11.25 | $33.75 |
| Gemini 2.5 Flash-Lite | $3.00 | $1.20 | $4.20 |
| Grok 4.7 | $60.00 | $18.00 | $78.00 |
| Grok 4.3 | $37.50 | $7.50 | $45.00 |
The lowest list price here is GPT-5 nano at $2.70, against $240 on GPT-5.5. Gemini rows use current prices and are for comparison; FUSR.ai does not route to Gemini.
Most summaries can wait, so batch them
If nobody is waiting on the screen, the batch APIs at Anthropic, OpenAI, and Google run the job within 24 hours at half the standard price. Caching helps less here because each document is different; only the shared instructions repeat.
| Model | List price | 10 in 100 input tokens cached | Batch price |
|---|---|---|---|
| Claude Sonnet 5.5 | $90.00 | $84.30 | $45.00 |
| Claude Haiku 5.5 | $4.50 | $4.23 | $2.25 |
| GPT-5.5 | $240 | $227 | $120 |
| GPT-6.1 Terra | $45.00 | $42.15 | $22.50 |
| GPT-6 Luna | $4.50 | $4.23 | $2.25 |
| GPT-5 mini | $13.50 | $12.83 | $6.75 |
| GPT-5 nano | $2.70 | $2.57 | $1.35 |
| Grok 4.7 | $78.00 | $73.50 | Not listed |
| Grok 4.3 | $45.00 | $41.85 | Not listed |
xAI does not list a batch price on its models page, so those rows say Not listed.
Check a cheap model is good enough
Small models summarize well, but they can drop a detail you care about. A quick test before you switch:
01Pick 30 real documents
Include the long ones and the messy ones, not only clean examples.
02Summarize them on your current model and the cheap one
Same prompt, same length limit.
03Read them side by side
Mark any summary that misses a fact you would need. If the cheap model misses none, switch.
Long documents cost more on some models
Haiku 5.5 charges more once a prompt passes 100K tokens, GPT-5.5 above 272K, and Grok and Gemini 3.1 Pro above 200K. Split very long documents into parts if you can. Full price lists are on LLM API pricing.