Choose LLM models and routing based on cost per feature
The agent tracks per-model spend for each AI feature across OpenAI, Anthropic, Bedrock and Vertex AI, and recommends where a cheaper model or a routing change could fit.
Runs in
Claude · Slack
First optimization recommendation
The next day
ROI
Pays for itself with the first recommendation you act on
For each AI feature, show last month's spend by model and provider, and the cost per 1,000 requests.
ticket-summary uses a frontier model for 2.1M requests/month averaging 600 tokens: €6,900. Routing requests under 1,000 tokens to a smaller model would cut that feature's cost by an estimated 70%, if your evals hold.
The AI team runs evals and decides on routing.
Playbook
What the agent does
From what Costory notices to what you decide. Costory suggests the next step in conversation, and you decide whether to go ahead.
Costory notices
A feature's model cost per request rises, or a new model ships.
Costory investigates
01
Allocate by feature
API keys, projects, IAM roles or request labels map LLM spend to features.
02
Break down by model
Per-model cost for each feature, across providers.
03
Add usage
Requests or tokens per feature from your own metrics.
04
Claude proposes
Routing or model changes with the expected cost impact; quality checks stay with your team.
Where you see it
Claude
You decide
The AI team runs evals and decides on routing.
Why it matters
Teams default to the biggest model for everything. Without per-feature, per-model cost, nobody knows where a smaller model would do the job.
Who it’s for
AI engineers, product engineers, CTOs
Where it runs
Claude · Slack
What you get
What lands in front of your team
Per-feature model costs
Know what each feature costs on each model.
Routing candidates
Where a smaller model could fit.
Your evals decide
Cost is one input; quality stays your call.
Sources and assistants
What it connects
Data sources
Example prompts
Or drive it yourself
Compare
For each AI feature, show last month's spend by model and provider, and the cost per 1,000 requests.
Recommend
Which features use a frontier model for short, simple requests? Suggest a routing change and estimate the impact.
FAQ
Questions about this playbook
Does Costory evaluate model quality?
No. Costory brings the cost side; model quality and evals stay with your team.
Which LLM providers are supported?
OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex AI and OpenRouter, plus AI media tools like ElevenLabs and Kling AI.
Does Costory guarantee savings?
No. The agent investigates and recommends with the evidence attached; your team decides what to change. Impact figures are estimates based on your own data.
Does Costory change anything in our cloud accounts?
No. Costory reads billing exports and the metrics you connect, and has no write access to your cloud. Actions it takes create Costory objects only: alerts, reports, dashboards and events. Your team stays in control of every change.
Step-by-step guide in the Costory docs
Related playbooks
Run these alongside it
Costory notices
Daily, as LLM provider costs refresh.
LLM cost allocation by team, feature and customer
Allocate model spend from OpenAI, Anthropic, Azure OpenAI, Bedrock, Vertex AI and OpenRouter to teams, features and customers.
See the playbook
Costory notices
Monthly, or when a feature's unit cost moves.
Cost per AI feature: tokens, GPUs and infrastructure in one number
Combine model API spend with the GPU and serving infrastructure behind each AI feature, then divide by usage for a true cost per request.
See the playbook
Costory notices
Every Monday, across every AI provider.
All AI spend in one view: OpenAI, Anthropic, Bedrock, Vertex AI and more
The agent tracks total AI spend across model APIs, cloud AI services and AI tools, and reports the weekly trend in one place.
See the playbook
Put this playbook on call
Included in every plan, from €250/month. 30-day free trial: connect in under 30 minutes, get your first optimization recommendation the next day.
First optimization recommendation
The next day
ROI
Pays for itself with the first recommendation you act on
Setup
About 30 minutes
Data residency
EU
Access
Read-only on your cloud