Amazon has closed Bedrock Agents Classic to new customers. Existing agents continue to run, and there's no announced end-of-life—but if you're planning a move to AgentCore, there's one migration step that deserves more attention:
Review your cost model before you migrate your code.
Your billing model has changed
With Bedrock Agents Classic, most teams only needed to think about:
- Model inference
- Lambda
- Knowledge Base queries
- Other AWS resources
The orchestration layer itself wasn't a separate billable service.
AgentCore changes that.
In addition to inference costs, you now have platform-level charges such as agent invocations and gateway requests. That means your total cost is no longer driven only by tokens.
Token tracking is no longer enough
Many internal cost dashboards look something like this:
interface SessionCost {
inputTokens: number;
outputTokens: number;
estimatedCost: number;
}
That worked well when inference represented almost the entire bill.
After migrating, your total session cost may also include platform activity that token-based tracking never sees.
The result isn't an incorrect calculation—it's an incomplete one.
Think in two cost layers
Instead of treating "AI cost" as a single number, separate it into two categories:
Inference costs
- Input tokens
- Output tokens
- Model pricing
Platform costs
- Agent invocations
- Gateway requests
- Other orchestration charges
Tracking both independently makes it much easier to understand why costs change over time.
A prompt optimization affects inference.
An orchestration redesign might barely change tokens while significantly increasing platform activity.
Those are different engineering problems.
Before migrating
A migration is a good opportunity to validate your assumptions.
I'd recommend comparing:
- Estimated inference cost
- Platform charges
- Combined session cost
for the same workload before rolling changes into production.
That baseline makes future regressions much easier to spot.
The broader trend
AgentCore isn't unique here.
Across the industry, agent platforms are becoming more opinionated and more capable—and that often means more of the bill moves into the orchestration layer instead of the model layer.
For years, tracking tokens was enough.
Going forward, teams will increasingly need to treat model costs and platform costs as separate budgets.
The sooner your observability reflects that split, the easier future migrations will be.
https://github.com/salimassili62-afk/ai-costguard
Top comments (0)