The cheapest way to summarise a million documents
When you summarise at scale you are buying input tokens in bulk. Output is almost free; context limits and rate limits are what really constrain you.
Summarisation inverts almost everything about chat economics. In chat, output length drives the bill. In bulk summarisation the output is a few hundred tokens against several thousand of input, so the input rate decides the cost, and the cheapest model by blended cost is often not the cheapest model for the job.
The workload used here
One million documents at 8,000 input tokens and 300 output tokens each — a corpus-scale run, not a demo. Ranked by input price, the cheapest options in the current dataset are:
| Model | Provider | Input /1M | Output /1M | Context | Cost for 1M docs |
|---|---|---|---|---|---|
| Qwen2.5 Coder 7B | Nebius | $0.01 | $0.03 | 33K | $89 |
| Qwen2.5 Coder 3B Instruct | Nscale | $0.01 | $0.03 | — | $89 |
| Qwen2.5 Coder 7B Instruct | Nscale | $0.01 | $0.03 | — | $89 |
| Nemotron 3.5 Lightning 30b A3b | Perplexity | $0.011 | $0.17 | — | $143 |
| Granite 4.0 H Micro | Cloudflare Workers AI | $0.017 | $0.112 | 131K | $169.6 |
| Mistral Nemo Instruct 2407 | DeepInfra | $0.019 | $0.03 | 131K | $161 |
When the document does not fit
The ranking above assumes a chunking strategy. If a single document is large enough that chunking loses the thread — a book, a contract bundle, a whole repository — the constraint becomes the context window, and the cheapest models in the market are frequently not available. Models with at least a 200K window price out differently:
| Model | Provider | Input /1M | Context | Cost for 1M docs |
|---|---|---|---|---|
| Databricks GPT 5 Nano | Databricks | $0.05 | 272K | $519.83 |
| Qwen Turbo | Alibaba DashScope | $0.05 | 1M | $460 |
| Qwen Turbo | Alibaba DashScope | $0.05 | 1M | $460 |
| Qwen Turbo | Alibaba DashScope | $0.05 | 1M | $460 |
| GPT 5 Nano | OpenAI | $0.05 | 272K | $520 |
Three things that move this bill more than the model choice
- Chunk overlap. A 10% overlap is a 10% surcharge on every run, and it is usually set out of caution rather than measurement.
- Prompt boilerplate per chunk. Instructions resent with every chunk multiply by the document count; batch jobs are the workload where caching and prompt trimming pay most.
- Retry policy. Malformed output that gets re-sent doubles the cost of that document. Schema validation before the retry is cheaper than the retry.
As with every figure on this site, these are published list rates normalised to US dollars per million tokens. Batch discounts, committed-use terms and free tiers are excluded, and a provider can change a rate at any time — the methodology page states the rules.
See every model ranked for input-heavy work, recomputed on each refresh.
Open the summarisation rankingFrequently asked
› Does the input/output price split matter for summarisation?
Yes, and it is the opposite of chat. A summary is short — a few hundred tokens against thousands of input tokens — so input price decides the bill and output price barely registers. Models that look expensive per blended token can be the cheapest choice for this shape of work.
› Do I need a long context window?
Only if a single document exceeds the model's window. For a corpus of ordinary documents you chunk and summarise, a 128K window is usually enough; long-context models matter when one document is a book, a legal bundle or a large codebase.
Where the numbers come from
The data pages this article draws on — check any figure against them, they are recomputed on every refresh.