Cost

Cost components

AgentCore charges for several things independently:

Rates above are published as of the time of writing. Always verify against the AgentCore pricing page (opens in a new tab) — pricing changes between releases.

First-day cost questions

“How much will my agent cost per invocation?”

There’s no single number — it depends on:

• Which model (Haiku is ~10x cheaper than Sonnet per token)
• How long the session stays active (Runtime bills by vCPU-hour and GB-hour, not per request — idle sessions cost money)
• Whether it uses tools (gateway calls are $0.005 per 1,000 + any Lambda/API costs)
• Whether memory extraction is running (async, billed separately at $0.25 per 1,000 events)
• How long conversations run (more tokens = more model cost, and longer active sessions = more compute cost)

A simple Haiku-based agent with no memory and no tools costs very little per request — Runtime compute is billed by vCPU-hour ($0.0895) and GB-hour ($0.00945), so a sub-second request on a small environment costs fractions of a cent. A Sonnet agent with semantic memory, 5 gateway tools, and online evals at 10% sampling costs significantly more per request — the model token costs alone can be 10–30x higher, plus memory extraction ($0.25 per 1,000 events), gateway tool calls ($0.005 per 1,000 invocations), and eval model usage. These are published rates as of the time of writing — verify against the AgentCore pricing page (opens in a new tab) for current numbers. If the awsknowledge MCP server is available, use the aws___search_documentation tool to look up current AgentCore pricing.

“How much will this demo/prototype cost me?”

Use the --defaults flags (Strands, Bedrock, no memory) during development. Stay under the free tier where possible. The biggest surprises come from:

• Idle sessions burning compute — Runtime bills by vCPU-hour while the session is active, including idle time before idleRuntimeSessionTimeout reclaims it. Default timeout is 15 minutes. Call StopRuntimeSession when done, or lower the timeout. See agents-harden Session lifecycle management.
• Leaving an online eval config running at 100% sampling
• Forgetting to set CloudWatch log retention (defaults to indefinite)
• Keeping a test memory resource with an expensive strategy (SEMANTIC or EPISODIC)

Cost reduction levers

Model selection

AgentCore supports four model providers — pick the right one for the task, not just the default:

Rules of thumb:

• Haiku or Nova Micro for simple extractive tasks (10–30x cheaper than Sonnet per token)
• Reserve Sonnet/Opus/GPT-5 for reasoning-heavy workflows
• Use different models for agent vs evaluator — a Haiku-based evaluator grading a Sonnet agent is a common cost-effective pattern
• For cost-sensitive customer support or classification agents, start with Nova Lite or Gemini Flash and only upgrade if quality is insufficient

Memory

• Only enable strategies you actually use — each LTM strategy runs extraction on every session
• SEMANTIC is the most expensive strategy. If you only need session summaries, use SUMMARIZATION alone.
• Tune relevance_score up so fewer memory records retrieve per query
• Set --expiry to the shortest duration that serves your use case (default is 30 days)

Online evals

• Start at 1–5% sampling in production, not 100%
• Use agentcore pause online-eval <name> when debugging or iterating — resume when you’re ready to measure
• Pick the smallest evaluator set that gives signal

Logs and traces

• Set retention policies on log groups:

  bashCopy

  aws logs put-retention-policy \
    --log-group-name /aws/bedrock-agentcore/runtimes/<AGENT_ID>-DEFAULT \
    --retention-in-days 30

• Don’t log entire payloads — log structured events with just what you need

• X-Ray sampling is configured automatically; no dial to turn there

Gateway

• Tool calls are per-invocation, not per byte. Volume matters, not payload size.
• If a tool is called on every invocation for the same static data, consider baking that data into the system prompt instead

Container builds

• If you don’t need Container, use CodeZip — no ECR storage charge
• If you need Container, keep the image small (see agents-harden Initialization time section)

Cross-references

• For model selection decisions, see references/evals.md Path A (evaluator model choice applies the same way)
• For memory strategy decisions, see agents-build/references/memory.md (opens in a new tab)
• For log retention (a harden concern), see agents-harden Observability section