Use case
Classify and tag content at scale
Classification is the highest-leverage use of an LLM: fixed labels, short inputs, clear evaluation. The main lever is the per-token price of a cheap profile.
Updated
Design the label set first
Fix the categories before you call the model. Fewer, mutually exclusive labels keep the output stable and easy to evaluate.
Ask for a single JSON tag, not free prose. A constrained output is cheaper to parse and simpler to verify.
Batch and budget
Classification over many documents is dominated by input tokens. Check the input price of the published profiles and pick the cheapest that still labels correctly.
Run a small labeled sample first, measure agreement with a known-good set, then scale the batch. A spend-limited key keeps a failed run bounded.
Cache the obvious, call the model for the rest
A hash of already-seen inputs avoids re-billing identical documents. Cached input price applies only when the provider reports cached tokens.
For bulk jobs, prefer non-streaming calls: they are simpler to retry and to reconcile against the journal.
Published model prices
USD per 1M tokens. The numbers come from the public catalog on every page rebuild — this is the sale price, not an estimate.
| Model | Input | Output | Cache |
|---|---|---|---|
| Claude Opus 5.5 | $2.70 | $8.00 | $0.14 |
| DeepSeek Flash | $0.50 | $2.00 | $0.01 |
| Claude Sonnet 5.5 | $1.33 | $6.67 | $0.07 |
| DeepSeek V4 Pro | $2.20 | $6.60 | $0.07 |
Top-up from $10, balance in USD. 1 USD = 1,000,000 UT. No subscription and no auto-charge.
Questions
Which profile is cheapest for classification?
The entry tier (flash) is priced lowest. Confirm the exact per-1M input price in the live table on /models.
Can I stream classification?
Chat streaming is supported, but a non-streaming call is simpler for bulk classification and reconciliation.
How do I know the batch cost in advance?
Estimate tokens per document, multiply by the catalog input price, then verify the total against /account/logs after the run.
Does the journal help with audits?
Yes: every request appears with model, token counts and the applied price for the last 30 days.