Firecrawl’s guide to cutting AI token costs starts with sending less

Firecrawl has published a nine-technique guide to reducing the token use and cost of AI applications. Its most useful point is that agents often keep paying to resend material they have already seen, so trimming retrieved content and structuring prompts can matter as much as choosing a cheaper model.

Firecrawl Watch analysis

What happened

The guide, published on 13 October, says an eight-turn agent can be billed repeatedly for the same page because model APIs resend conversation history on each turn. It recommends measuring input, output and cached tokens by task before making changes, then targeting the largest source of usage.

Firecrawl’s practical examples include asking a tool for a specific answer instead of returning a whole page, using clean markdown when a model needs the full document, and requesting structured fields when the answer can be checked by code. It also recommends keeping stable prompt content at the start of requests to improve cache reuse. Read Firecrawl’s guide.

Key findings

• Measure usage before optimising
  Track input, output and cache tokens per task to find where the bill is growing.
• Send only the content needed
  Firecrawl says one stock-price answer used 6 characters, compared with 295 characters from Claude Code’s fetch tool.
• Treat search output as a cost choice
  On an Openbenchmarks search-only board, Firecrawl used a median 7,456 tokens per task and completed 70.3%; Perplexity used 8,765 tokens and completed 77.3%.
• Make prompts cache-friendly
  Stable content at the start of a prompt can help providers reuse it instead of billing it as fresh input.

Why it matters

For anyone building an AI agent, token costs are shaped by what gets sent, how often it gets repeated and whether the provider can cache it. A shorter result can reduce spending, but the benchmark figures also show why token count alone is a poor trophy: Perplexity used more tokens than Firecrawl in that comparison, yet completed more tasks.

Firecrawl presents its guide as advice for developers building applications and agents on the OpenAI or Anthropic APIs. Its own examples and cited benchmarks are useful starting points, not a guarantee that the same savings will hold for every model or workload.

Our read

Start by logging token use per task, then cut the biggest input source and check that the output still does the job. That is a better first move than optimising by vibes, or discovering that a very cheap agent has become an exceptionally affordable way to be wrong.

What to watch

• Whether the proposed reductions hold up on workloads beyond the guide’s examples.
• Whether lower token use changes task completion or answer quality.
• How much savings come from fewer tokens versus cheaper cached or batched pricing.

Discussion spark: When optimising an AI agent, should teams prioritise the lowest token cost per task, or pay more for a tool that completes more tasks reliably?

not affiliated with or endorsed by Firecrawl