“How much will this cost?” is the hardest question to answer about AI automation. Token pricing is per-million, models charge differently for input and output, and the cost of a single run depends on how much text goes in and comes out. Most teams either ignore costs until the bill arrives or over-optimize by using the cheapest model for everything.
Neither approach works. Here’s how to think about AI workflow costs practically.
The basics: what a single recipe run costs
A recipe run has two cost components: input tokens (your prompt + context) and output tokens (the AI’s response). Output tokens are typically 3-5x more expensive than input tokens.
Here’s what typical recipe runs cost across providers (approximate, as of early 2026):
Simple extraction or classification (short input, structured output):
- Claude Haiku 4.5: ~$0.002-0.005 per run
- GPT-5-mini: ~$0.002-0.004 per run
- Gemini 2.5 Flash Lite: ~$0.001-0.003 per run
Content generation (moderate input, longer output):
- Claude Sonnet 4.5: ~$0.01-0.03 per run
- GPT-5.1: ~$0.01-0.025 per run
- Gemini 2.5 Pro: ~$0.008-0.02 per run
Complex analysis with extended thinking (long input, reasoning + output):
- Claude Opus 4.5: ~$0.05-0.15 per run
- o3: ~$0.04-0.12 per run
- Gemini 3 Pro: ~$0.03-0.10 per run
These are rough ranges. Actual costs depend on input length, output length, and whether features like web search or extended thinking are enabled.
What a workflow costs
A workflow chains multiple recipes. The cost is the sum of all step costs. Here’s a realistic example:
Invoice Processing Workflow (4 steps):
- Extract invoice data (Haiku 4.5): $0.003
- Check discrepancies (Sonnet 4.5): $0.015
- Generate approval summary (Haiku 4.5): $0.002
- Write final report (Sonnet 4.5): $0.02
Total per run: ~$0.04
Run this 50 times a week (10 invoices per day), and the monthly cost is about $8. For context, the manual process takes a finance team member 15-20 minutes per invoice — roughly 12 hours of work per week.
New Lead Pipeline (4 steps with web search):
- Prospect research with web search (Sonnet 4.5): $0.04
- Lead qualification (Haiku 4.5): $0.005
- Condition check: free (logic only, no AI call)
- Draft outreach (Sonnet 4.5): $0.02
Total per run: ~$0.065
At 200 leads per month, that’s about $13/month in AI costs.
The optimization lever: per-step model selection
The biggest cost optimization in JieGou isn’t using cheaper models everywhere — it’s using the right model for each step.
In the invoice workflow above, extraction and the approval summary use Haiku (fast, cheap, good at structured tasks). The discrepancy check and final report use Sonnet (better reasoning, better prose). If you used Opus for everything, the workflow would cost ~$0.35 per run instead of $0.04 — nearly 10x more for marginal quality improvement on the simpler steps.
JieGou lets you set the model independently for each recipe, so you can optimize at the step level without changing the workflow structure.
Estimating costs before you run
JieGou includes a cost estimator that projects the cost of a workflow run before you execute it. The estimator uses historical token usage from previous runs of each recipe to predict costs for the current run.
For new recipes with no history, the estimator uses model-specific defaults based on the input and output schema sizes.
Tracking costs after you run
The analytics dashboard breaks down token usage and costs by:
- Recipe — Which recipes are most expensive per run?
- Workflow — What’s the total cost of each workflow?
- Department — How much is each team spending?
- Model — What’s the cost distribution across providers?
This visibility lets you spot optimization opportunities. If one recipe accounts for 60% of a workflow’s cost, that’s where to experiment with a different model or a shorter prompt.
The monthly cost picture
For a team running moderate automation:
| Workload | Monthly runs | Estimated cost |
|---|---|---|
| Sales lead pipeline | 200 | $13 |
| Marketing content repurposing | 20 | $1.50 |
| Support ticket triage | 800 | $4 |
| Weekly deal review | 4 | $0.80 |
| Invoice processing | 200 | $8 |
Total: ~$27/month in AI provider costs. That’s the actual token cost, paid directly to the providers. JieGou’s platform subscription is separate and doesn’t include AI markup.
Compare that to the manual time replaced — dozens of hours per week across the team — and the ROI is hard to argue with.
Rules of thumb
- Use the cheapest model that produces acceptable output. For triage, extraction, and classification, Haiku or GPT-5-mini is usually sufficient. Save Opus and o3 for complex analysis and high-stakes content.
- Shorter prompts cost less. A concise prompt template that tells the AI exactly what to do costs less than a long one padded with examples and caveats.
- Structured output schemas reduce waste. When the AI knows exactly what fields to fill, it produces shorter, more focused output. Less output = lower cost.
- Web search adds cost. The search itself is included in the model’s pricing, but the search results add to the input context. Only enable web search for recipes that need current information.