Ask HN: How are people forecasting AI API costs for agent workflows?

We hit this exact problem and built NORNR (nornr.com) around it. Instead of trying to forecast costs upfront, we enforce them at runtime. Every agent action that involves spend has to request a mandate first. Policy decides approved, queued, or blocked based on budget, rate, context, whatever you define. Every decision gets a signed receipt so you have a full audit trail. It doesn't replace your cost estimates but it puts a hard ceiling on what agents can actually spend. Free tier if you want to try it.

We've been dealing with this exact problem building agent-driven workflows. A few things that have actually helped:

The unpredictability is worse than the absolute cost. Our billing model broke several times not because costs were high, but because we couldn't bound them. One approach that helped: define a 'token budget' per user action at design time - cap total tokens per session and treat hitting the cap as a first-class outcome your product handles gracefully, not an error.

On the forecasting side, we track cost per workflow step rather than per request. Step-level cost is much more stable than request-level because it absorbs the variance in tool calls and retries. Once you have step costs, you can forecast by expected workflow composition.

On fixed subscription pricing for AI APIs - I'd actually pay a premium for that. The unpredictability creates a hidden cost: you over-provision margins and add complexity to your pricing tier design. A flat rate for a capacity bucket would eliminate both.

The question I'd ask about any such service: how do they handle the tail cases where agents go off-rails and rack up 10x normal token usage? That's where the cost risk actually lives.

I’d break it into a few separate numbers instead of trying to estimate one blended API cost.

We usually look at cost per workflow run, runs per active account, and the heavier paths separately, then keep retries and tool calls as their own line items. That makes the pricing side easier to reason about.

this is exactly the problem we keep running into. the cost isn't just "how many tokens did this call use," its "how many tokens did this entire user action consume across all the agent loops, retries, tool calls, and embeddings."

most observability tools show you the LLM call as one flat span. you can see it cost X tokens but you cant correlate it with the API request that triggered it, or see that the agent looped 4 times because the first 3 outputs failed validation. so you end up building custom logging and hoping the numbers add up.

we've been building an APM (immersivefusion.com) where cost is a first-class dimension on every trace. so you can see one request flow from the UI through your backend through the agent workflow, and each span carries its token cost. the idea is you should be able to answer "what does a checkout cost when the recommendation agent is in the loop" without stitching together 3 different tools.

for the forecasting question specifically, i think the answer is you need a few weeks of production data with good instrumentation and then you can build a distribution. the variance is real but its not random, its usually a few specific flows that blow up (retries on bad structured output like @hkonte mentioned, or RAG queries that hit the wrong chunk size). once you can see which flows are expensive the guardrails become obvious.

also wrote a longer piece on this if anyone's interested: immersivefusion.com/blog/end-to-end-observability-from-ui-to-ai-agent-to-invoice

I love the idea. @Edgee.ai we are tracking cost in real time by tag, LLM, ...but not yet on forecast, and indeed it will be very useful. Something to explore; thanks for the feedback.

Just add very hard high limits and add instrumentation so you can track it and re-evaluate it accordingly.

This takes a couple of hours maximum at best.

It feels like the traditional fixed SaaS pricing model is slowly shifting toward more consumption-based pricing.

Honestly, if you're designing your agent workflows properly with hard limits on retries and tool calls, the variance shouldn't be that wild. Most of the unpredictability comes from not having those guardrails in place early on. A few weeks of real production data usually shows the average cost is more stable than you'd expect.

imo switch to local models could be an option