AWS made its next-generation Amazon Bedrock AgentCore Runtime the default for newly created runtimes in 19 AWS Regions on November 12, 2026. Existing runtimes stay on their current version until their owners move them, according to the official announcement.
The new platform, V2, restores sessions from a snapshot of an initialized agent, adjusts compute and memory to the workload, and scales idle agents to zero. AWS reports startup times below two seconds at the 75th percentile for a 2 GB container image, compared with 29.7 seconds on the previous runtime.
The billing change deserves equal attention. AWS claims up to 68% less billed memory per session for multi-turn agents. That is a reduction in memory usage charged to the customer, not a promise that the total bill falls by 68%. Published secondary reporting lists higher V2 unit rates, making the balance between active compute and reclaimed memory central to the migration decision.
What Changes for New and Existing Runtimes
AgentCore Runtime is AWS’s managed execution environment for agents. Each session continues to run inside its own hardware-isolated Firecracker microVM; V2 changes how that environment starts and scales.
AWS’s announcement sets out these capabilities:
| Capability | What AWS says V2 supports |
|---|---|
| Session compute | Up to 8 vCPU |
| Session memory | Up to 32 GB |
| Container image size | Up to 10 GB |
| Session startup | Under two seconds at P75 for a 2 GB image |
| Idle behavior | Scale to zero, with memory reclaimed within minutes of release |
| Multi-turn memory efficiency | Up to 68% less billed memory per session |
The default applies across five Regions in the Americas, seven in Europe and seven in Asia Pacific. These include all three listed US Regions, Canada Central, São Paulo, Frankfurt, Ireland, London, Tokyo, Singapore and Sydney. AWS provides the complete Region list in its announcement.
Three deployment rules matter:
- New runtime, no version specified: AWS says it uses V2 in the listed default Regions.
- Existing runtime: It stays on its current platform version until explicitly moved.
- Explicit version selection: Set
platformVersiontoV2when creating or updating a runtime to select the new platform. Set it toV1when creating a runtime that should remain on the older platform.
The announcement also directs builders to select V2 explicitly in Regions where it is not yet the default.
I’d pin the platform version in deployment configuration even when choosing V2. A default change can otherwise make a newly created environment behave differently from an older environment built from the same unversioned configuration.
Faster Starts Need a Precise Definition
AWS’s startup mechanism moves initialization work ahead of session creation: V2 restores an initialized agent snapshot instead of starting every session through the previous process. The practical benefit is less waiting before an agent can begin work.
P75 describes the 75th percentile of measured starts. It leaves the slowest quarter outside the headline figure. It also measures session startup, not necessarily the time until a model finishes its response, a browser completes navigation or a compilation ends.
The AWS technical blog makes a broader claim that sessions start in under two seconds independently of image size and concurrency. The November announcement’s concrete comparison specifies a 2 GB image and P75.
Small independent measurements point in a similar direction, but with much narrower evidence:
| Published measurement | Image size | V1 start | V2 start |
|---|---|---|---|
| AWS announcement, P75 | 2 GB | 29.7 seconds | Under 2 seconds |
| Classmethod test, as reported by 9to5Amazon | 46 MB | 3.4 seconds | 2.1 seconds |
| Classmethod test, as reported by 9to5Amazon | 859 MB | 13.6 seconds | 1.8 seconds |
9to5Amazon’s November 16 analysis attributes the independent figures to Classmethod engineer suzuki.ryo, testing in Tokyo. Each figure was a single measurement, so those observations cannot establish a percentile or behavior under sustained concurrency.
That report also describes a slower first warm call on V2, 1.912 seconds versus 0.611 seconds, and runtime creation or update taking about three minutes instead of roughly five seconds. These are different operations from the advertised session-start measurement.
The useful implication is to separate deployment readiness, new-session startup and subsequent invocation latency in migration testing. An improved cold-start number can coexist with a slower deployment or a different first-call result.
Lower Billed Memory Can Still Mean a Higher Bill
AWS says idle agents scale to zero and memory is reclaimed within minutes of release. Elastic memory helps most when an agent’s footprint rises during active work and falls during pauses between turns. A continuously busy agent with a persistent memory footprint has less idle capacity to reclaim.
The reclaim delay matters, too. A conversational pause lasting only a few seconds offers a different savings opportunity from a session left inactive for several minutes.
For AI pricing, the key calculation is:
Compute cost = billed vCPU-hours × CPU rate + billed GB-hours × memory rate.
9to5Amazon reports these US-dollar rates from the AgentCore pricing page:
| Resource | V1 | V2 consumption pricing |
|---|---|---|
| CPU, per vCPU-hour | $0.0895 | $0.1276 |
| Memory, per GB-hour | $0.00945 | $0.0169 |
These are rates reported by the publication, not a fresh verification of AWS’s live pricing page. Confirm applicable pricing before budgeting. The comparison below uses consumption pricing and excludes committed-baseline options.
Calculated from those rates, V2 CPU is about 43% more expensive per vCPU-hour, and memory about 79% more expensive per GB-hour. Lower usage must offset those increases.
For memory alone, V2 breaks even when billed GB-hours fall by approximately 44.1%:
1 − ($0.00945 ÷ $0.0169) ≈ 44.1%
That threshold says nothing about the CPU portion of the bill.
A Worked Example: 68% Less Memory, About 1% Less Cost
Consider an illustrative session with these assumptions:
- The session lasts 20 minutes.
- V1 bills a 1 GB memory footprint throughout that period.
- Both versions consume two minutes of billed CPU at 1 vCPU.
- V2 reduces billed memory by the percentages shown below.
- Model inference, external tools, storage and network charges are excluded.
This is arithmetic using the reported rates, not a measured AgentCore workload.
| Scenario | CPU cost | Memory cost | Total compute cost | Change from V1 |
|---|---|---|---|---|
| V1 baseline | $0.002983 | $0.003150 | $0.006133 | — |
| V2, 30% less billed memory | $0.004253 | $0.003943 | $0.008197 | +33.6% |
| V2, 50% less billed memory |
Frequently Asked Questions
3 questions
1Will Existing AgentCore Runtimes Move to V2 Automatically?
No. AWS says existing runtimes stay on their current version until moved. Select platformVersion: V2 when creating or updating a runtime to choose V2.