# AWS Bedrock Adds OpenAI Agents, Cuts Costs 22% [2026] > Source: > Published: 2026-10-12 20:12:02+00:00 Amazon Web Services spent the first two weeks of October 2026 doing something it rarely does in public: admitting that its own agent platform had become too expensive to run blind. On October 7, AWS quietly updated the Bedrock documentation history to note that cost allocation by IAM principal, meaning the ability to tie a dollar figure to the specific user or role that triggered it, now extends to the newer bedrock-mantle inference endpoint, not just the original bedrock-runtime path. Three weeks earlier, AWS had opened a public preview of Bedrock Managed Agents powered by OpenAI, built on a customized version of OpenAI’s Agents API and wired directly into AWS-native resources. Alongside it came the AWS Well-Architected Agent, a service that reviews a customer’s own AWS environment and hands back fixes. And by the end of October, AWS plans to turn on a committed-baseline pricing tier for Bedrock AgentCore Runtime that cuts the published on-demand compute rate by roughly 22 percent for teams willing to commit to a baseline. None of these four moves made headlines on their own. Together, they describe a cloud vendor racing to solve a problem it helped create: AI agents that run for minutes or hours, call tools, hold memory across turns, and generate bills nobody can trace back to a team, a project, or a reason. Amazon Bedrock AgentCore pricing has become searchable in its own right because finance teams are typing it into Google at 1,600 searches a month in the US, according to keyword volume data pulled in October 2026, a number that tracks almost exactly with the panic building inside FinOps teams over agent spend they cannot explain. ## What AWS actually shipped this quarter Three separate AWS releases converged inside a few weeks of each other, and it is worth separating them because the [AWS announcements blog](https://aws.amazon.com/blogs/machine-learning/category/post-types/announcements/) has treated them as one undifferentiated “AWS ships more AI agent stuff” story. The first is Bedrock Managed Agents powered by OpenAI, a public preview that lets a team run OpenAI’s own Agents API inside Bedrock rather than stitching together a self-hosted runtime. AWS built the integration to be AWS-native, meaning the agent reads and writes to the customer’s own AWS resources, Identity and Access Management roles included, rather than operating as an external black box that happens to call AWS APIs over the network. The second is the AWS Well-Architected Agent, also in public preview, which analyzes a customer’s existing AWS environment and produces recommendations across cost, security, performance, and resilience, aligned to whatever business goals the organization has stated. The idea is straightforward: instead of waiting for a human solutions architect to run a Well-Architected review once a year, a persistent agent watches the account continuously and surfaces ready-to-implement fixes as problems appear, rather than after a quarterly audit. The third, and the one with the most immediate financial bite, is the IAM-principal cost allocation change. Amazon Bedrock now supports attributing inference spend to the specific IAM user or role that generated it, for requests made through the bedrock-mantle endpoint, extending a capability that previously only covered bedrock-runtime. In plain terms, a company running dozens of agents across different teams can finally see which role, which service account, or which individual triggered a given charge, instead of reconciling CloudTrail activity logs against a lump-sum bill after the fact. ## The pricing mechanics behind AgentCore AgentCore is the runtime layer underneath most of this, and AWS has published granular, consumption-based pricing for it with no upfront commitment or minimum fee. The numbers matter because they explain why cost allocation suddenly became urgent. On v1 microVMs, AWS billed CPU at $0.0895 per vCPU-hour and memory at $0.00945 per GB-hour. The v2 generation, which is the current default execution environment, charges $0.1276 per vCPU-hour and $0.0169 per GB-hour on demand, a straightforward increase tied to the architecture’s expanded capability. What changes by the end of October 2026 is the arrival of a committed-baseline rate sitting alongside the on-demand price: $0.0997 per vCPU-hour and $0.0132 per GB-hour for v2 microVMs, for teams that commit to a baseline level of usage rather than paying pure consumption rates. Do the arithmetic and the discount lands at just under 22 percent on both CPU and memory, a meaningful cut for any organization running agents at real production volume rather than isolated pilot projects. AWS’s own documentation walks through a worked example that illustrates how quickly these line items stack up: a 600-second session consuming 4 GB of memory at the v1 rate works out to roughly $0.0063 in memory cost alone, before CPU, networking, or any NAT gateway charges that keep billing while an agent sits idle. Runtime instances are charged against the underlying EC2 session cost plus a 12 percent management fee in AWS’s documented example, and idle memory is only reclaimed after 120 seconds, a detail that matters for any agent architecture that pauses between tool calls rather than terminating cleanly. Memory itself carries its own fee schedule separate from compute. Short-term memory costs $0.25 per 1,000 new events. Long-term memory storage using AWS’s built-in memory strategies runs $0.75 per 1,000 stored records per month, while overwrite or self-managed strategies drop that to $0.25 per 1,000 records monthly. Retrieving long-term memory costs another $0.50 per 1,000 retrievals. Each of those is individually small. Multiplied across thousands of concurrent agent sessions inside a large enterprise, they are exactly the kind of charge that used to vanish into an undifferentiated AWS bill until someone asked the CFO why cloud spend jumped without a corresponding revenue signal. ## Why cost attribution became the urgent problem Shared AI infrastructure has a specific accounting flaw that traditional compute never had at the same scale: a single model endpoint can serve ten different teams, twenty different agents, and a few hundred human users simultaneously, and the resulting invoice line simply says “Bedrock inference” with a dollar total. Before IAM-principal attribution, a finance team trying to figure out which project burned through its quarterly AI budget had to manually cross-reference CloudTrail activity logs against the billing data, a process that does not scale past a handful of agents and becomes functionally impossible once an organization is running agents that spawn sub-agents. Industry FinOps commentary circulating in late September and early October 2026, including discussion on the [What’s New in Cloud FinOps podcast](https://podcasts.apple.com/au/podcast/whats-new-in-cloud-finops/id1534914412), has been blunt about where this is heading. That discussion has centered on four specific controls that organizations are now demanding: spending caps per agent, approval workflows before an agent can escalate its own resource usage, hard per-agent budgets that cut off execution rather than just alerting after the fact, and unified cost aggregation that rolls agent spend into the same dashboards finance teams already use for traditional cloud costs. That list reads less like a product roadmap and more like a checklist written by people who have already been burned by a runaway agent loop that nobody noticed until the monthly bill arrived. The container layer underneath these agents is shifting in parallel. AWS has been pairing agent workloads with Amazon ECS Express Mode and Amazon EKS Auto Mode, and promoting the EKS Provisioned Control Plane with a 99.99 percent availability SLA, a detail that signals containers remain the default substrate for running AI agents in production rather than a transitional stepping stone toward something agent-specific. For an infrastructure team deciding where to place a new fleet of agents, that reliability number carries real weight when the agents themselves are making decisions that touch production systems. ## How Bedrock’s governance push compares to Microsoft and Google AWS is not acting in isolation. Microsoft’s Foundry Agent Service, the managed agent offering inside Azure AI Foundry, takes a structurally different approach to the same cost problem. According to [Microsoft’s published pricing page](https://azure.microsoft.com/en-us/pricing/details/foundry-agent-service/), there is no additional charge for creating or running Foundry-native agents that use prompts and workflows, meaning the orchestration layer itself is free. Hosted agents, by contrast, are billed according to the underlying container compute they consume per hour, a model that pushes the cost conversation toward infrastructure metering rather than per-agent-session billing. That is a meaningfully different philosophy from AWS’s approach, where AgentCore charges for CPU, memory, short-term memory events, long-term memory storage, and long-term memory retrieval as five separate line items. Microsoft’s model is simpler on paper, free orchestration plus metered hosted compute, but it also means a Foundry customer loses the granular per-component attribution that AWS is now building directly into IAM-level billing. Which approach wins out probably depends on whether a given enterprise cares more about predictable billing simplicity or granular chargeback precision, and large regulated enterprises with internal cost-center requirements tend to prefer the latter even when it adds billing complexity. The practical effect is that AWS, Microsoft, and Google are now competing on governance tooling almost as much as they compete on raw model access. An enterprise buyer evaluating agent platforms in October 2026 is no longer just asking which vendor has the best model catalog. They are asking which vendor can tell them, in real time, which internal team is responsible for a cost spike, and which vendor will let them cap an agent’s spend before it happens rather than explain it after the invoice lands. | Capability | AWS Bedrock AgentCore | Microsoft Foundry Agent Service | |---|---|---| | Orchestration/agent creation cost | Billed as part of CPU and memory consumption | No additional charge for Foundry-native prompt and workflow agents | | Compute billing unit | $0.1276 per vCPU-hour, $0.0169 per GB-hour (v2 on-demand) | Underlying container compute, billed hourly | | Committed-rate discount | ~22% off CPU and memory, launching by October 2026 | Not published in current pricing material | | Memory/session pricing | Separate short-term and long-term memory fees per 1,000 events/records | Not broken out as a separate line item | | Cost attribution | IAM-principal attribution now covers bedrock-mantle and bedrock-runtime | Not established in published pricing documentation | ## The historical arc: from AgentCore GA to committed pricing AgentCore’s trajectory this year has been a steady climb from experimental runtime to billable production infrastructure with governance bolted on after the fact, which is roughly the opposite order most mature cloud services follow. AgentCore first reached general availability earlier in 2026 as AWS’s answer to a specific gap: agents that call tools, hold memory across turns, and run for minutes or hours do not fit the request-response model that Lambda was originally built around. The original v1 runtime worked, but cold starts were rough, with AWS’s own published figures putting v1 cold-start latency between 5.4 and 30 seconds depending on container image size. AWS closed much of that performance gap with AgentCore Runtime v2 in September 2026, pushing p75 cold-start latency down to roughly 1.9 to 2.0 seconds for container images between 200MB and 2GB, a drop of around 90 percent off the high end of the old range. That same release window also brought GPT-6.1 Sol onto Bedrock for reasoning, coding, computer-use, and professional workloads, alongside xAI’s Grok 4.7 for coding and long-running agent scenarios, and expanded availability of Anthropic’s Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5 into new regions through Bedrock’s cross-Region inference routing. What October adds to that timeline is the governance layer that was conspicuously missing when the performance story shipped in September. Cost allocation by IAM principal, a committed-baseline pricing tier, a Well-Architected reviewer agent, and an OpenAI-powered managed agent option all landed within roughly three to four weeks of each other. That sequencing, fast first, cheap and accountable second, mirrors how AWS has rolled out other infrastructure categories historically. Lambda itself launched with minimal cost visibility tooling and only grew detailed per-function cost breakdowns years later, once customers had already built enough serverless infrastructure that bill shock became a recurring support ticket rather than an edge case. ## What this means for enterprise AI budgets For a platform team currently running production agents on Bedrock, the practical to-do list from this round of releases is fairly concrete. First, the committed-baseline pricing tier is worth evaluating against actual usage patterns before it launches broadly by the end of October, since a roughly 22 percent discount on both CPU and memory is large enough to change the economics of an agent fleet that currently runs at steady, predictable volume rather than bursty, unpredictable spikes. Second, IAM-principal cost allocation is not automatic just because AWS turned the capability on. Teams still need to structure their IAM roles so that distinct agents, projects, or business units map to distinct principals, otherwise the new attribution data collapses back into the same undifferentiated lump sum it was meant to replace. Organizations that have historically run everything through a single shared service role will need to do real identity-architecture work before this feature delivers the chargeback clarity it promises. Third, the arrival of Bedrock Managed Agents powered by OpenAI gives procurement teams a genuine alternative path: instead of choosing between a fully AWS-native agent stack and a fully external OpenAI-hosted deployment, a team can now run OpenAI’s Agents API inside AWS’s own security and billing boundary. That matters most for regulated industries where data residency and IAM integration are hard requirements that previously ruled out third-party agent platforms regardless of model quality. ## The idle-cost problem nobody budgeted for One detail buried in AWS’s own AgentCore documentation deserves more attention than it has gotten: CPU charges do not apply during model or tool I/O wait, provided no background process is using CPU during that window, but memory remains billable for the entire duration a session stays alive. That asymmetry means an agent that spends most of its session waiting on a slow downstream API call, a common pattern for agents that chain multiple tool calls together, racks up memory charges the entire time even though it is doing essentially no compute work. Combined with the 120-second window before idle memory gets reclaimed, this creates a specific failure mode: agent architectures that pause frequently between steps, rather than running continuously or terminating cleanly, pay a meaningful memory tax for behavior that looks efficient from a compute-utilization standpoint but is not efficient from a billing standpoint. Engineering teams optimizing agent code for latency and compute efficiency may be solving the wrong problem if their actual bill is being driven by sessions that sit open waiting on external responses. | AgentCore cost component | Rate | What triggers it | |---|---|---| | v2 CPU (on-demand) | $0.1276 per vCPU-hour | Active compute during agent reasoning/execution | | v2 CPU (committed baseline) | $0.0997 per vCPU-hour | Committed usage tier, launching by October 2026 | | v2 memory (on-demand) | $0.0169 per GB-hour | Billable for entire session duration, including I/O wait | | v2 memory (committed baseline) | $0.0132 per GB-hour | Committed usage tier, launching by October 2026 | | Short-term memory events | $0.25 per 1,000 new events | Each new memory event the agent writes | | Long-term memory storage (built-in strategy) | $0.75 per 1,000 records/month | Records retained under AWS’s managed memory strategies | | Long-term memory retrieval | $0.50 per 1,000 retrievals | Each time the agent reads back stored memory | ## Competitive pressure from the open-source agent framework angle AWS has also been pushing an open-source agent framework that, notably, runs on rival clouds rather than locking customers exclusively into Bedrock. That is an unusual move for a hyperscaler, since it theoretically weakens AWS’s own platform lock-in. The likely logic is that AWS would rather be the default orchestration layer that developers reach for, even if some of those workloads eventually execute on Azure or [Google Cloud](https://shattered.io/google-cloud-finops-flexible-gpu-commitments-focus-1-4-2026/) compute, than lose developer mindshare entirely to a framework built by a competitor or a neutral open-source foundation. That strategic calculation sits alongside the governance push described above. A cloud vendor that controls both the orchestration framework and the cost-attribution tooling developers rely on is in a strong position to remain the default choice even when the underlying compute could technically run anywhere. Cost allocation by IAM principal, specifically, only works cleanly within AWS’s own identity system, which gives AWS a retention lever that has nothing to do with model quality or raw compute price and everything to do with how deeply a customer’s financial reporting has become dependent on AWS-native tooling. ## Predictions for the next two quarters Several trends from this round of releases, which [cloud industry coverage](https://www.cloudcomputing-news.net/) has flagged as part of a broader cost-governance shift, look likely to continue through early 2027. First, expect per-agent spending caps and approval workflows to move from a FinOps wish list into an actual shipped AWS feature within the next two quarters, given how explicitly the committed-baseline pricing tier and IAM-principal attribution are already pointing in that direction. Second, expect Microsoft and Google to respond with their own granular cost-attribution tooling for Foundry Agent Service and Vertex AI respectively, since ceding that governance ground to AWS would be a competitive mistake once enterprise buyers start treating cost transparency as a baseline requirement rather than a differentiator. Third, the committed-baseline discount on AgentCore pricing will likely be the first of several tiered pricing options AWS introduces for agent compute, following the same pattern AWS used for EC2 Reserved Instances and Savings Plans, where an initial simple discount tier eventually fragments into multiple commitment lengths and discount depths. Fourth, expect Bedrock Managed Agents powered by OpenAI to graduate from public preview to general availability within two to three quarters if adoption numbers hold up, following the typical AWS preview-to-GA cadence for high-interest services. Fifth, and most speculatively, expect the first publicly reported case of a company disclosing a specific runaway agent-cost incident in an earnings call or SEC filing sometime in 2027, once enough enterprises have scaled agent deployments to the point where a configuration mistake can produce a material line item rather than a rounding error. ## What budget-conscious teams should do now Teams running or planning agent deployments on Bedrock have a narrow window to get ahead of these changes rather than reacting to them after the committed-baseline tier and expanded cost allocation are fully rolled out. Auditing current IAM role structure to ensure agents map to distinct, traceable principals is the single highest-leverage action available right now, since every other governance feature AWS has shipped this quarter depends on that mapping already being in place. Modeling actual usage against the committed-baseline discount before committing is the second priority, because a roughly 22 percent savings only materializes if usage patterns are steady enough to justify a baseline commitment rather than pure on-demand consumption. For agents that run sporadically or in bursts, the on-demand rate may still be cheaper in practice despite the lower headline rate on the committed tier. Finally, teams evaluating whether to adopt Bedrock Managed Agents powered by OpenAI should weigh the AWS-native integration benefits, meaning direct access to IAM, VPC, and AWS resource permissions, against the fact that this remains a public preview without long-term pricing or SLA commitments. Preview services are the right place to experiment and build internal expertise, but production-critical workloads with strict uptime requirements may want to wait for general availability before fully committing. ## Where Google Cloud fits into the agent cost conversation Google Cloud is pursuing the same agent infrastructure race through Vertex AI’s agent tooling, though its published pricing and governance documentation for agent-specific cost attribution is less granular in public materials than what AWS and Microsoft have now disclosed. That gap does not mean Google lacks the capability internally, but it does mean enterprise buyers currently comparing all three hyperscalers on cost governance alone have clearer, apples-to-apples pricing data for AWS and Microsoft than for Google’s equivalent agent stack. Expect that disclosure gap to close quickly, since none of the three hyperscalers can afford to look less transparent than its rivals once procurement teams start asking cost-attribution questions during vendor evaluations. The broader pattern across all three clouds is that agent cost governance has gone from an afterthought to a genuine purchasing criterion in under a year. A platform team choosing between Bedrock, Foundry Agent Service, and Vertex AI in October 2026 is weighing model quality, runtime performance, and now cost transparency as three roughly equal factors, a shift that would have seemed premature twelve months ago when most agent deployments were still pilot projects rather than production systems with real budgets attached. ## Frequently asked questions **What is Amazon Bedrock AgentCore?** AgentCore is AWS’s set of managed capabilities for deploying and operating AI agents, covering Runtime, Gateway, Identity, Memory, Observability, Browser Tool, and Code Interpreter components. Each component can be used independently, billed on a consumption basis with no upfront commitment or minimum fee. **How much does Bedrock AgentCore cost?** On v2 microVMs, on-demand pricing is $0.1276 per vCPU-hour and $0.0169 per GB-hour. A committed-baseline tier launching by October 2026 cuts those rates to $0.0997 per vCPU-hour and $0.0132 per GB-hour, a discount of roughly 22 percent for teams that commit to a baseline usage level. **What is Bedrock Managed Agents powered by OpenAI?** It is a public preview AWS feature that runs a customized version of OpenAI’s Agents API natively inside Bedrock, integrated with the customer’s own AWS resources and IAM permissions, rather than requiring a separate externally hosted OpenAI agent deployment. **What does the AWS Well-Architected Agent do?** It is a public preview service that analyzes a customer’s AWS environment and generates recommendations across cost, security, performance, and resilience, aligned with the organization’s stated business goals, with ready-to-implement fixes rather than just a diagnostic report. **What changed with Bedrock cost allocation?** AWS extended IAM-principal cost allocation, the ability to attribute inference spend to a specific user or role, to the bedrock-mantle endpoint, a capability that previously only covered the bedrock-runtime endpoint. This lets organizations trace agent spend back to specific teams or service accounts inside their Cost and Usage Report. **How does this compare to Microsoft’s Foundry Agent Service?** Microsoft does not charge extra for creating or running Foundry-native agents built on prompts and workflows, billing only for underlying container compute on hosted agents. AWS instead charges for CPU, memory, and multiple memory sub-components separately, trading billing simplicity for more granular cost attribution. **Why does memory keep billing even when an agent is idle?** AWS’s documentation confirms that CPU charges do not apply during model or tool I/O wait when no background process is active, but memory remains billable for the full duration a session stays alive, and idle memory is only reclaimed after 120 seconds. **Is Bedrock Managed Agents powered by OpenAI ready for production use?** It is currently in public preview, which typically means pricing, SLAs, and feature scope can still change before general availability. Teams with strict uptime or compliance requirements should treat it as an evaluation and prototyping option rather than a production-ready commitment until AWS moves it to GA.