
AWS deal with Greece GettlyImages/ LIONEL BONAVENTURE
Amazon Web Services' rebuilt AgentCore runtime charges 79% more per gigabyte-hour of memory than the original, so whether an AI agent's hosting bill falls depends on how long that memory sits idle. AWS detailed the next-generation runtime on Thursday, Nov. 12, pairing sessions of up to 8 vCPUs and 32 GB of memory with billing that follows memory actually in use.
In AWS's own pricing example, a customer-support agent handling 1 million 10-minute sessions a month costs $6,703 on the new runtime. That example gives no side-by-side figure against the original runtime. FactualMinds, an AWS Select Tier Services Partner, priced the same session profile at original-runtime rates, with peak memory held for the full session, and got about $5,429 a month, which makes the new runtime roughly 23% more expensive for that workload. That figure is a model, not an AWS invoice, but its inputs are AWS's published rates.
AWS argues the higher rate applies to far fewer gigabyte-hours because memory is loaded on demand and reclaimed. AWS's Nov. 9 weekly roundup says the team expects most agents to see a lower bill, though that is not guaranteed.
Published consumption rates for the new runtime are $0.1276 per vCPU-hour and $0.0169 per GB-hour, against $0.0895 and $0.00945 for the original. Those are increases of about 43% for CPU and 79% for memory.
| Runtime | CPU per vCPU-hour | Memory per GB-hour |
|---|---|---|
| Original (v1) | $0.0895 | $0.00945 |
| New (v2), consumption | $0.1276 | $0.0169 |
| New (v2), committed baseline | $0.0997 | $0.0132 |
Billing runs per second with a 1-second minimum and a 128 MB memory minimum. CPU scales to zero while an agent waits on a model, tool or database call, and the pricing page says idle memory is reclaimed automatically after 120 seconds.
The math behind the disagreement is simple. FactualMinds' break-even math assumes the original runtime bills peak memory for the whole session. Because memory costs 79% more per hour, the billed average must fall to about 56% of that peak (0.00945 divided by 0.0169) before memory spending breaks even. AWS's own example shows why that bar is high: its agent's memory climbs from 1 GB to 2 GB to 2.5 GB over 10 minutes and does not fall, so the average sits near 65% of the peak.
FactualMinds modeled two other cases. Fifty thousand five-second chat turns a month came to $6.87 on the original runtime and $10.03 on the new one, about 46% more, because turns end long before the 120-second window opens. In the opposite case, 120 eight-hour sessions that spike to 2.5 GB and then fall to a 128 MB floor came to $31.27 on the original runtime and $14.89 on the new one, about 52% less. The firm labels both as modeled list-rate math, not client bills.
Other analysts reach a different verdict for short sessions. Sarvar Nadaf, an AWS Community Builder, wrote on Oct. 5 that short, bursty sessions that free memory between steps favor the new runtime, while long sessions holding a steady, large working set favor the original; his post says the opinions are his own and may not reflect AWS's. Rajesh Beri of The D*AI*LY Brief modeled 300,000 four-turn chat sessions on Oct. 31 and concluded the new runtime is "usually the right choice despite its higher sticker price." His model is also hypothetical, not measured invoices.
The sources leave one gap. The pricing page ties reclaim to 120 seconds of idleness, while AWS's September post also describes reclaiming memory when an agent releases it, such as per-request buffers, without publishing figures for how much that returns. Teams can settle the question only by measuring their own memory profile.
Instead of booting and initializing each new instance, the runtime prepares an agent's environment once, captures a snapshot of the running container after it reports healthy, and restores that snapshot for every new instance. One-time work such as loading model artifacts is baked into the snapshot, and AWS trims the snapshot to the working state needed to resume, leaving out caches and transient memory so its size stays roughly flat as images grow.
In AWS's September benchmark, 75th-percentile cold starts held at about 2 seconds (1.9 to 2.0) for container images from 200 MB to 2 GB. The original runtime climbed from about 5.4 seconds to nearly 30 over the same range. AWS now says starts are up to 15x faster and need no warm pools.
The September figures suggest where the 15x may come from. Dividing nearly 30 seconds by about 2 gives 15, which matches the 2 GB image; at the 200 MB image, 5.4 seconds against about 2 is a gain of roughly 2.7 times. The post does not define its baseline for the claim. Three more limits apply. The benchmark used an empty echo agent that called no model or tools, and the agent's own code took about 34 milliseconds, so almost all measured time was platform start. The timings are client-side and include a cross-region round trip. And the benchmark stopped at 2 GB, one-fifth of the 10 GB image ceiling AWS now advertises, so the flat-start claim for larger images is untested in the published data. AWS itself notes that model calls, not startup, dominate what users wait on in production.
Faster starts move the wait to deployment. Creating or updating a runtime on the new platform takes several minutes while the snapshot is prepared, and the container must report healthy within 120 seconds. Developers opt in by setting the platformVersion field to V2. AWS's release notes add an OverheadLatency metric for every V2 runtime that isolates platform time from agent processing, which lets teams check the cold-start claim in their own accounts.
The technique that speeds starts also turns every instance into a clone. AWS's developer guide says every restored instance starts from an identical copy of the snapshot's memory, including the state of random number generators in cryptographic libraries and any keys, tokens or identifiers created during startup.
The guide's instructions follow from that. Secrets, session IDs and random values must be generated inside the request handler, not at startup, and hostname and process ID cannot serve as unique identifiers because they are identical across instances. Firecracker's own documentation describes the underlying hazard: restored clones resume with the same random number generator state, and the document says it is unclear whether the timer and hardware noise mixed in is enough for security, so it recommends reseeding after each restore. An agent that signs requests or mints tokens with a value created at startup would share it with every other instance restored from the same snapshot.
AWS says each session still runs in its own hardware-isolated Firecracker microVM with its own identity, policy and audit trail. The clone risk sits above that boundary, in application code, so the isolation claim does not remove it.
For always-on agents, AWS is adding committed baseline pricing: a per-session memory baseline billed as a minimum, with on-demand bursts up to 4x that baseline. The pricing page lists committed rates of $0.0997 per vCPU-hour and $0.0132 per GB-hour, about 22% below the new runtime's consumption rates but still about 11% above the original runtime's CPU rate and about 40% above its memory rate.
AWS says steady workloads are billed at Lambda MicroVM rates, and the numbers line up. Lambda's published Arm rates in US East (N. Virginia) of $0.0000276944 per vCPU-second and $0.0000036667 per GB-second work out to $0.0997 per vCPU-hour and $0.0132 per GB-hour. The post itself gives no dollar figures.
Timing is murkier. The pricing page lists the committed rates as launching by October 2026, yet AWS's Nov. 9 roundup still calls committed baseline pricing "coming soon." The sources reviewed do not say whether those rates are billable today.
AWS's post says the runtime is available across regions in North America, South America, Europe and Asia Pacific, without naming them. The new runtime reached general availability on Sept. 18 in five regions: us-east-1, us-east-2, us-west-2, eu-west-1 and ap-northeast-1. The developer guide's release notes say a November update added Frankfurt, Stockholm, Mumbai, Singapore and Sydney, bringing the total to 10 regions, with no South American region named. Teams planning a rollout, especially in South America, should confirm their region before committing.
The same release notes do not list the 8 vCPU, 32 GB or 10 GB image limits that the post announces, so within the sources reviewed those figures rest on AWS's blog alone.
Agents can now mount a filesystem, such as an Amazon S3 bucket exposed through FUSE, and save work to it, and AWS says that data stays inside the session's trust boundary under the same controls as the session. FUSE, short for Filesystem in Userspace, lets a program in user space present storage such as an object bucket as ordinary files, which is why a bucket can look like a local folder to an agent.
Per-session storage scoping, which would bind each session to its own folder through session-labeled credentials, is not yet available. AWS says it will arrive by re:Invent, which is scheduled for Nov. 30 to Dec. 4 in Las Vegas. The release notes separately describe session context keys that let sessions sharing an execution role be limited to their own data under one policy, but they do not say whether that covers mounted storage.
AWS cited customer results, all self-reported. Coinbase said cold starts fell from almost a minute to under five seconds. Mission Cloud reported a 17% reduction in measured runtime cost per call, in what AWS describes as an initial comparison of the same workload. Zuora said the runtime cut memory footprint per invocation by around 60% in many use cases. Mission Cloud's figure is the only customer cost number among the statements reviewed.
Okta, Deloitte, Caylent, Rescale and Cox Automotive also supplied statements without figures. Okta said the new runtime and Lambda MicroVMs let it safely scale autonomous engineering agents.
By FactualMinds' modeling, agents with long sessions that spike in memory and then sit idle are the clearest candidates, and agents with short, steady sessions are the clearest risk. FactualMinds frames the case for switching around cold-start consistency, not guaranteed savings, and notes that the public AgentCore pricing calculator still uses original-runtime peak-memory rates dated July 4. The same analysis says CloudFormation and the AWS Cloud Development Kit cannot yet set the platform version.
AWS's September post suggests hiding start time for interactive agents by opening a session as soon as a user starts a chat, so the environment warms while the user types.
The new AgentCore runtime pairs faster, flatter cold starts with a higher price per unit, and the trade favors agents whose memory spikes and then idles. Teams should compare their billed average memory against 56% of their old peak before switching, a threshold that assumes the original runtime bills peak memory. They should treat AWS's up-to-15x and up-to-68% claims as best cases, since the 15x baseline is undefined and the 68% figure covers memory per session for multi-turn agents with natural pauses. Teams should also confirm region and committed-rate availability and move secrets out of startup code. AWS's answer on per-session storage scoping is due when re:Invent opens on Nov. 30.
When billed average memory falls below about 56% of the original runtime's peak, assuming the original bills peak memory, since the memory rate is 79% higher. In FactualMinds' modeled scenarios, long sessions that spike and then idle came out cheaper on the new runtime, while five-second chat turns came out about 46% higher.
Set the platformVersion field to V2 when you create or update a runtime. FactualMinds says the original runtime stays the default if the field is omitted at creation.
AWS's release notes put the new runtime in 10 regions, five from the Sept. 18 launch plus Frankfurt, Stockholm, Mumbai, Singapore and Sydney, and name no South American region.
Every instance restores from the same memory snapshot, so any key, token or random value created at startup is identical across instances. AWS's guide tells developers to generate such values inside the request handler.
