benchr
What this guide covers
This guide brings together three practical cost questions. The price-per-use-case table compares representative workloads across major commercial models. The context-window article explains why a published limit is not always a usable limit. The long-context analysis compares full-context requests with retrieval-based approaches.
Pricing by workload
-
The price-per-use-case table
A cost comparison for six workloads and three frontier models. In the article's scenarios, chat costs $0.014 per turn on Sonnet, RAG queries cost $0.036, document summaries cost $0.18, and an uncapped agent session can exceed $50.
-
AI API price history
Providers rewrite pricing pages and old numbers vanish. benchr's append-only log keeps every verified price, change, and sunset — dated, source-linked, and downloadable as open data under CC BY 4.0.
Context-window economics
-
Context windows compared, across four frontier models
Compare advertised capacity with reproducible retrieval probes on your own documents. The article's 600K and 800K points are labeled editorial hypotheses, not provider limits or measured guarantees.
-
The million-token context was always a marketing number
In the article's example, a 200K-token Claude Opus request costs about $1, while the retrieval-based version costs $0.06. That is a 17× difference per question. Test both approaches with your own documents before choosing an architecture.
When to skip the frontier entirely
-
RAG vs fine-tuning, with the math
A conditional decision guide: compare data freshness, retrieval failure, behavior change, review cost, and dated scenario assumptions before choosing RAG, fine-tuning, or both.
-
Small language models: a local deployment test plan
A reproducible way to compare classification accuracy, failure cost, hardware, and reviewed output cost. No unpublished 1,200-email result is presented as evidence.
The cost discipline that works
Three controls can make production AI spending easier to predict.
One: constrain output. Cap max-tokens. Force structured formats. Instruct "no preamble" and trim everywhere. Output is where the money goes. See the prompt-engineering piece for the techniques.
Two: cache repeated prefixes. Anthropic, Google, and OpenAI support prompt caching, with some cached-input rates near 10% of standard input pricing. If the same system prompt appears on every call, compare the cache rules and effective rate for your chosen model.
Three: route by workload. A small local model may suit classification, while a hosted mid-tier model may suit routine generation. Reserve more expensive models for tasks where your evaluations show a useful gain. The comparison tool helps estimate the trade-offs.