# Amazon Bedrock AgentCore (Runtime / Code Interpreter / Browser)

Official pricing: https://aws.amazon.com/bedrock/agentcore/pricing/  
Category: hyperscaler · Isolation: firecracker

## Pricing regimes (raw)

- **Code Interpreter session (container; active CPU + peak memory)** (resource): $0.0895/vCPU-h, $0.00945/GiB-h
- **Runtime v1 microVM (same meter as Code Interpreter)** (resource): $0.0895/vCPU-h, $0.00945/GiB-h
- **Runtime v2 consumption (idle memory reclaimed after 120 s)** (resource): $0.1276/vCPU-h, $0.0169/GiB-h
- **Runtime v2 committed baseline (announced, launching by Oct 2026)** (resource): $0.0997/vCPU-h, $0.0132/GiB-h [sales, beta]
- **Browser Tool session (1 vCPU / 4 GB)** (resource): $0.0895/vCPU-h, $0.00945/GiB-h [alt]
- **Runtime Instances (EC2 in your account + 12% management fee)** (sizes): c7g.xlarge 4 vCPU/8 GiB $0.1624/h; m7g.xlarge 4 vCPU/16 GiB $0.182784/h; c7i.xlarge 4 vCPU/8 GiB $0.19992/h; m7i.xlarge 4 vCPU/16 GiB $0.225792/h; c7g.2xlarge 8 vCPU/16 GiB $0.3248/h; c7i.2xlarge 8 vCPU/16 GiB $0.39984/h [alt]

## Features

Yes: persistent disk, volumes, idle auto-stop, custom image (Docker or snapshot), your own Docker/OCI image, full VM (own kernel), Docker inside, root, browser, computer-use API, code interpreter, browser automation API, desktop control API, browser + desktop control, egress allowlist, open internet, static egress IP, private networking, BYOC, SOC 2, HIPAA, SSO, GPU, arm64, Python, Node.js, wake on request, audit logs, EU data residency, spend limits, MCP server, hosted agent API (their own agent), gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), extra volumes, shared volumes

No: snapshots, memory snapshots, fork/clone, pause/resume, ≥24 h sessions, start from your own snapshot, nested virtualization, systemd, desktop GUI, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, SSH, public IPv4, HTTPS preview URLs, custom domains, raw TCP inbound, self-hosting, open source, Windows, macOS, GPU desktop, HTTP method/path egress rules, live resize, live fork (no pause), memory fork, webhooks, automatic snapshots, snapshots on demand, agent harness API, secret proxy, secret proxy for any API

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/aws-agentcore.json → feature_evidence

## Caveats

- Runtime v2 committed baseline ($0.0997/vCPU-h + $0.0132/GB-h) is announced as 'launching by October 2026'; term, minimum and purchase mechanics unpublished, so it is opt-in only (flag sales).
- Rates verified in the AWS Price List API (AmazonBedrockAgentCore, publication 2026-09-15) and the pricing page (2026-09-28).
- Active-CPU billing: cpu_basis active with no floor; I/O wait (LLM calls) is free only if no background process burns CPU. Memory: Code Interpreter, Browser and Runtime v1 use ram_basis peak (peak consumed up to that second, i.e. peak-to-date in the session), so the workload's peak RAM input drives the bill; Runtime v2 reclaims idle memory after 120 s and uses ram_basis active. The old peak_mem_frac knob was removed (the engine's peak RAM input replaces it). 128 MB memory minimum not encoded.
- Runtime v2 committed baseline is announced only ("launching by October 2026"); mechanics unknown, flagged beta.
- A single session maxes at 2 vCPU/8 GB (Browser 1 vCPU/4 GB), so a 4 vCPU workload does not fit consumption modes.
- The pricing page Instances example cites "c7g.2xlarge $0.867 x 0.12"; c7g.2xlarge on-demand is $0.29 in us-east-1 and the API fee is $0.0348 (= 12% of $0.29). The page example looks wrong.
- Other AgentCore meters (not modelled): Gateway $0.005/1k invocations, search $0.025/1k, $0.0002 per tool-index-month; Memory short-term $0.25/1k events, long-term $0.75/1k records-month (built-in) or $0.25 (custom), retrieval $0.50/1k; Web Search tool $7/1k queries; Knowledge Base $1/1k retrievals; Evaluations per token.
- Storage: direct-code deployments at S3 Standard rates; container images in ECR. No snapshots/pause for consumption sessions.
- Network: "standard EC2 data transfer rates"; Internet egress (AWS Price List API, AWSDataTransfer us-east-1, 2026-09): first 100 GB/month free aggregated across all services and regions, then $0.09/GB first 10 TB, $0.085 next 40 TB, $0.07 next 100 TB, $0.05 over 150 TB. eu-west-1 is also $0.09; ap-south-1 $0.1093; ap-northeast-1 $0.114; ap-southeast-1 $0.12 (first-10-TB tier). The engine uses the flat us-east-1 first tier.
- AWS Free Tier $200 credit for new accounts is generic; no AgentCore-specific free quota.
- Runtime v1 (microVM, custom container image) and Code Interpreter (containerized, AWS-managed images) share the same v1 meter but are separate modes (runtime-v1 / code-interpreter). Runtime v2 reclaims idle memory at higher unit rates.
- Idle tail: a Runtime microVM session stays alive (and keeps billing memory, CPU ~0) until idleRuntimeSessionTimeout, default 900 s, unless the caller stops it; Code Interpreter sessions default to a 900 s timeout too. For short sessions this tail dominates: a 1-minute call at 2 GB bills ~16 min of memory. Engine assumes billing ends when work ends.
- Runtime session storage (preview): a per-session filesystem of up to 1 GB on microVM Runtime modes persists across stop/resume (14-day idle expiry, reset on version update). It is free during the public preview; the future price is unannounced. Not available for Code Interpreter or Browser.
- Correction: the pricing-page Instances example is per 3-hour session (3 h x $0.289/h c7g.2xlarge = $0.867; fee $0.867 x 0.12 = $0.104). It is consistent apart from $0.289 vs the $0.29/h list price. Supersedes the caveat calling the example wrong.

## How this provider charges


AgentCore's consumption regimes bill **active CPU** (CPU scales to zero during I/O wait such as LLM calls) and
**peak memory consumed up to that second** (128 MB minimum), per second with a 1-second minimum. Session hardware is
fixed at 2 vCPU / 8 GB (Browser 1 vCPU / 4 GB). Rates verified in the **AWS Price List API** (AmazonBedrockAgentCore
offer, publication 2026-09-15) and on https://aws.amazon.com/bedrock/agentcore/pricing/ (2026-09-28).

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Code Interpreter | Managed Python/JS/TS sandbox sessions | Active vCPU-s + peak GB-s, 1 s min, 128 MB memory floor | $0.0895/vCPU-h, $0.00945/GB-h | Price List API CodeInterpreter:Consumption-based:* |
| Browser Tool | Managed headless browser sessions (1 vCPU/4 GB, 10 GB disk) | Same | $0.0895 / $0.00945; profiles at S3 Standard rates | Price List API BrowserTool:* |
| Runtime v1 microVM | Agent runtime sessions (custom container or direct code) | Same; memory **not reclaimed while idle** | $0.0895 / $0.00945 | Price List API Runtime:Consumption-based:vCPU/Memory |
| Runtime v2 consumption | New runtime generation | Same; **idle memory reclaimed after 120 s** | $0.1276 / $0.0169 (+43% / +79%) | Price List API Runtime:*-v2 (us-east-1, us-east-2, us-west-2, eu-west-1, ap-northeast-1 only) |
| Runtime v2 committed baseline | **Announced**, "launching by October 2026" | Committed baseline; mechanics unpublished | $0.0997 / $0.0132 | pricing page (not in the API yet) |
| Runtime Instances | EC2 capacity in your account (sessions persist up to 14 days; up to 20 agents per capacity-provider session) | EC2 on-demand (billed by EC2) + management fee per instance-hour | 12% of EC2 on-demand (c7i.xlarge $0.02142/h, c7g.xlarge $0.0174, m7i.xlarge $0.024192); G-series GPU 7.8% (g6.xlarge $0.062774) | Price List API Runtime:Instance-based:*:Management-Hours |
| Regions | 14 commercial regions (us-east-1/2, us-west-2, ca-central-1, eu-central-1, eu-north-1, eu-west-1/2/3, ap-northeast-1/2, ap-south-1, ap-southeast-1/2) | Identical rates | $0.0895 / $0.00945 everywhere; GovCloud us-gov-west-1 $0.1074 / $0.01134 (+20%) | Price List API (18 region files) |
| Gateway | MCP/tool gateway | Per invocation / per indexed tool | $0.005 per 1k invocations, $0.025 per 1k searches, $0.0002 per tool-index-month; VPC target data $0.006/GB | Price List API |
| Memory | Agent memory | Per event / record / retrieval | Short-term $0.25/1k events; long-term $0.75/1k records-month built-in ($0.25 custom); retrieval $0.50/1k | Price List API |
| Web Search tool | Built-in search | Per query | $0.007/query | Price List API |
| Knowledge Base | Retrieval | Per query / storage | $0.001/query ($0.004 agentic); $5/GB-month | Price List API |
| Storage / network | Direct-code deployments; container images; egress | S3 Standard; ECR; "standard EC2 data transfer rates" | 100 GB/month free then $0.09/GB | pricing page, AWSDataTransfer |
| Free tier | New AWS accounts | Credits | Up to $200 (generic AWS; no AgentCore quota) | earlier research |

## Gotchas

1. **Active-CPU is real**: an agent waiting on an LLM 70% of the time pays ~30% of the CPU price — but any background process (a watcher, a dev server) keeps CPU "active".
2. **Memory is peak-to-date**, not average; one spike to 6 GB bills 6 GB for the rest of the session (v1). v2 reclaims idle memory after 120 s but costs 79% more per GB.
3. **Hard 2 vCPU / 8 GB per session** (Browser 1/4). Bigger workloads need Runtime Instances (EC2 + 12%).
4. **No pause/snapshot** for consumption sessions; 15-minute idle timeout ends the session (adjustable), 8 h max lifetime; Runtime session storage is capped at 1 GB.
5. **Committed baseline is not purchasable yet**; its rates equal Lambda MicroVMs Graviton rates.
6. The pricing page's Instances example ("c7g.2xlarge $0.867 x 0.12") contradicts the API fee ($0.0348 = 12% of $0.29).

## Worked example

4 vCPU / 8 GiB doesn't fit one session, so model **2 sessions of 2 vCPU / 4 GB** per unit: 100 concurrent
sessions x 176 h = **17,600 session-hours**, 30% CPU active, peak memory = 4 GB (conservative). 50 GiB of state
cannot be retained (no snapshots; S3 persistence not priced). 100 GiB egress inside the free tier.

| Regime | Per session-hour | Total / month |
|---|---|---|
| Code Interpreter / Runtime v1 | 2 x 0.3 x 0.0895 + 4 x 0.00945 = $0.0915 | **$1,610.40** |
| Same, peak memory only 1 GB | 0.0537 + 0.00945 = $0.06315 | **$1,111.44** |
| Runtime v2 consumption (memory never idle-reclaimed) | 2 x 0.3 x 0.1276 + 4 x 0.0169 = $0.14416 | **$2,537.22** |
| Runtime v2 committed (allocated 2 vCPU/4 GB, working hours only) | 2 x 0.0997 + 4 x 0.0132 = $0.2522 | **$4,438.72** (if the commit must cover 24/7: 100 x 730 x 0.2522 = $18,410.60) |
| Runtime Instances, 50 x c7i.xlarge (4/8) | 0.1785 + 0.02142 = $0.19992 per instance-hour x 8,800 | **$1,759.30** (+ EBS, IPv4 $44) |