OpenAI’s AWS partnership makes stateful AI a cloud-control-plane battle

Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.

OpenAI has not simply launched a finished “stateful AI” product on AWS. On February 27, 2026, OpenAI and Amazon announced a jointly developed Stateful Runtime Environment for agents. OpenAI models and Codex subsequently became generally available through Amazon Bedrock, while some managed-agent capabilities remained in preview.

The strategic importance is broader than model access. AWS is positioning itself to control the identity, networking, permissions, observability, compute, billing, and workflow infrastructure surrounding OpenAI-powered agents. That could move power in enterprise AI from the model API toward the cloud control plane—but companies should not assume the full runtime is generally available, separately priced, or ready for every production workload.

#	Preview	Product	Price
1		Nimo AI NAS, Agentic Computer Mini PC and AI Server, AMD Ryzen 7 PRO 8845HS(up to 5.1 GHZ, beat...	$1,999.99	Buy on Amazon

The launch timeline matters

The February partnership announcement, the later Bedrock model releases, and the proposed stateful runtime are related but distinct developments.

The February announcement describes the strategic partnership and Stateful Runtime Environment. It does not, by itself, establish that the runtime was generally available. As of the cited August 18 reporting point, the safest description is that the environment was announced and being developed, while the model and Codex offerings had reached GA.

#1 Best Overall

Nimo AI NAS, Agentic Computer Mini PC and AI Server, AMD Ryzen 7 PRO 8845HS(up to 5.1 GHZ, beat i5-1235u) up to 132TB ZFS Hybrid Storage, Dual 10GbE for 24hr AI Agent

• [Local AI Inference & 70B Model Ready] Equipped with the AMD Ryzen 7 PRO 8845HS processor, NEXUS is engineered for heavy local AI workloads. With a full-size GPU bay, it runs 70B LLMs natively without an internet connection. Ideal for AI developers and tech enthusiasts who need private environment for coding and model testing.
• [132TB Mass Storage with ZFS Integrity] Features a hybrid storage architecture (3×NVMe + 4×3.5" HDD) supporting up to 132TB. Utilizing the enterprise-grade ZFS file system and ECC memory, it prevents data corruption and bit rot—a must-have for professional photographers and video editors safeguarding 4K/8K RAW footage.
• [OpenClaw-Driven Automation Workflow] The built-in OpenClaw execution layer allows complex automated tasks to be processed locally. Even when offline, your backup schedules and AI file organization continue seamlessly. Say goodbye to monthly cloud subscriptions and high latency.
• [Dual 10GbE & USB4 Ultra-Connectivity] Experience server-class speeds with dual 10GbE ports and a 40Gbps USB4 interface. It enables multi-user real-time collaboration on large project files directly from the NAS, ensuring zero-lag editing for creative studios and production teams.
• [Open-Source ZimaOS for Total Privacy] Running on the fully open-source ZimaOS, NEXUS ensures your data stays physically on-premise with no backdoors. It acts as a "Digital Fortress" for privacy-conscious families and small businesses who demand absolute data sovereignty.

$1,999.99

Buy on Amazon

What “stateful AI” means here

Stateful AI is not merely a chatbot that remembers earlier messages. In this context, statefulness refers to the execution environment maintaining structured information across a multi-step task.

That can include:

• Conversation and working context
• Memory and prior work
• Tool-call history
• Workflow progress
• Compute access
• Identity and permission boundaries
• Long-running or resumable execution
• Operational recovery and audit records

A conventional model API returns an answer to a request. The application developer typically has to build the surrounding system: store context, decide which tools may run, handle retries, resume interrupted jobs, enforce permissions, and record what happened.

OpenAI’s proposed runtime is intended to absorb more of that orchestration burden. The Stateful Runtime announcement describes an environment where agents can preserve context and workflow state while accessing tools, compute, and governed resources.

Stateful does not mean autonomous, infallible, permanently self-aware, or capable of remembering everything. State can be bounded, configured, permissioned, deleted, or lost when a workflow ends. Human approval, authorization design, evaluation, and cost controls remain necessary.

What is available on Bedrock now?

OpenAI models

OpenAI models and Codex became generally available on Amazon Bedrock on June 1, 2026. AWS says customers can call them through an OpenAI-compatible Responses API, with usage priced at parity with OpenAI’s first-party rates and potentially eligible to count toward AWS commitments. See the AWS GA announcement for the named offering and conditions.

On July 13, AWS announced GA availability for GPT-5.6 Sol, Terra, and Luna. The models run through the bedrock-mantle endpoint and inherit AWS deployment controls such as IAM, VPC operation, and CloudTrail logging, subject to the product’s documented behavior.

The cited regions were:

• GPT-5.6 Sol: US East (N. Virginia) and US East (Ohio)
• GPT-5.6 Terra: US East (N. Virginia), US East (Ohio), and US West (Oregon)
• GPT-5.6 Luna: US East (N. Virginia), US East (Ohio), and US West (Oregon)

Regions, quotas, supported features, and service tiers change. Verify the AWS availability announcement and model documentation before committing a regulated or geographically constrained workload.

Codex and managed agents

Codex is OpenAI’s coding agent and is available through Bedrock access paths including the Codex CLI, desktop app, and Visual Studio Code extension. AWS credentials and Bedrock infrastructure are used for access. The April announcement described Codex and Bedrock Managed Agents as limited preview; the June announcement established GA for the named model and Codex offering.

Amazon Bedrock Managed Agents powered by OpenAI are a separate part of the story. AWS described agents with individual identities, action logging, and execution in the customer’s environment with inference through Bedrock. That is related to the broader stateful-runtime direction, but it should not automatically be treated as identical to every capability promised in the February Stateful Runtime announcement.

Frontier

OpenAI Frontier is an enterprise platform for building, deploying, and managing teams of AI agents. AWS was described as its exclusive third-party cloud distribution provider. That does not make AWS the exclusive cloud provider for all OpenAI workloads, and it does not mean OpenAI is abandoning Microsoft Azure.

How developers can access an OpenAI model through Bedrock

AWS documents the OpenAI-compatible endpoint format as:

https://bedrock-mantle.{region}.api.aws/openai/v1

For US East (N. Virginia), a minimal Python pattern is:

from openai import OpenAI

client = OpenAI(
    api_key="AWS_BEARER_TOKEN_BEDROCK",
    base_url="https://bedrock-mantle.us-east-1.api.aws/openai/v1",
)

response = client.responses.create(
    model="openai.gpt-5.6-terra",
    input="Summarize the latest project status."
)

print(response.output_text)

The cited model IDs are openai.gpt-5.6-sol, openai.gpt-5.6-terra, and openai.gpt-5.6-luna. AWS’s Bedrock Mantle documentation says the Responses API supports stateful conversation management, streaming, background processing, multi-turn interactions, and references to earlier turns through previous_response_id.

This code is illustrative, not a guarantee that every model supports every parameter or tool. Check the model-specific documentation for authentication, supported features, quotas, regional availability, and endpoint behavior.

Why this is a control-plane shift

The important change is not just that an OpenAI model can be called from an AWS account. Bedrock and related AWS services can place the surrounding agent system inside the enterprise’s existing operational framework.

• Identity: IAM can define who or what an agent is allowed to access.
• Networking: VPC-oriented architecture and PrivateLink can help connect workloads to private systems, subject to service-specific design.
• Audit: CloudTrail and related logging can record actions and administrative events.
• Policy: Guardrails and permission controls can constrain model-driven behavior.
• Data gravity: AWS-hosted databases, storage, applications, and internal services are already part of the environment.
• Procurement: Bedrock usage may align with existing AWS commitments and billing processes.

The AWS April announcement describes Bedrock integration with IAM, private networking, guardrails, encryption, CloudTrail, and other controls. The practical effect is that AWS can become the operational substrate for the agent even when OpenAI supplies the model intelligence.

This creates a possible inversion of the AI stack:

1. The model provider supplies reasoning and generation.
2. The cloud provider supplies execution, identity, policy, memory, observability, and billing.
3. The enterprise adopts the cloud-native runtime as the system through which agents act on business systems.

If an agent’s state, permissions, tool definitions, logs, compute, and data access all live in AWS, changing models may be easier than changing the control plane. That is an analytical implication of the partnership—not an announced claim that AWS owns OpenAI’s entire control plane.

Why OpenAI wants AWS

OpenAI gains a route into enterprises that already buy, govern, and operate through AWS. That can reduce the friction of introducing OpenAI capabilities into companies concerned with procurement, network architecture, auditability, and internal data access.

The partnership also combines distribution with infrastructure. OpenAI said Amazon would invest $50 billion, beginning with $15 billion and a further $35 billion subject to conditions. The announcement also described an expansion of an existing $38 billion multi-year agreement by $100 billion over eight years and approximately two gigawatts of AWS Trainium capacity. These are commitments described by the companies, not independently validated operating results.

For OpenAI, the relationship offers:

• AWS enterprise distribution
• Access to customers with large AWS commitments
• A route for Frontier distribution
• AWS infrastructure and Trainium capacity
• A managed path for production agents

For AWS, the deal offers more reasons for customers to keep AI workloads, data, and agent operations within AWS. It also strengthens AWS’s position against Microsoft Azure and Google Cloud in a market where the valuable layer may be enterprise execution rather than raw model hosting.

The competitive map

The strategic question is therefore not only “Which model scores highest?” It is also “Which platform governs the agent’s state and actions, and how portable is that governance?”

What using Bedrock changes compared with OpenAI direct

“Running in your AWS environment” should not be read as self-hosting OpenAI’s model weights or owning the inference hardware. It more accurately describes AWS-controlled application placement, identity, networking, data access, logging, and supporting infrastructure, with inference delivered through Bedrock.

Nor does Bedrock automatically mean that every request stays inside a customer’s private network or that every OpenAI feature behaves identically on both platforms. Check retention, data handling, endpoint behavior, regions, quotas, and feature parity for the exact service.

Pricing: token rates are only the beginning

AWS’s cited Bedrock pricing page listed these on-demand US East prices:

AWS announced on July 30 that Luna prices were reduced by 80% and Terra prices by 20%, while Sol pricing remained unchanged. Prices vary by region and service tier, so use the current Bedrock pricing page rather than treating these figures as permanent.

Total agent cost can also include:

• AgentCore runtime charges
• Tool and external API calls
• Storage and memory
• Retrieval and embeddings
• Network transfer
• Logging and observability
• Provisioned or reserved capacity
• Human-review operations
• Failed, duplicated, or repeated tool calls

Statefulness may reduce repeated context transmission or engineering effort, but it is not established as universally cheaper. Persistent memory can increase storage, retrieval, logging, and execution costs.

Risks enterprises should examine before committing

State lock-in

Persistent state is valuable because it is structured and durable. That can also make migration difficult. Before relying on a vendor-managed runtime, ask whether state can be exported, whether its schema is customer-controlled, whether another model can resume a workflow, and whether tool contracts, logs, and traces remain portable.

A larger security surface

A stateful agent may retain prompts, tool results, identity references, business context, customer data, and approval history. Review retention periods, encryption, tenant isolation, access policies, deletion procedures, legal holds, residency, and the possibility that prompt injection could persist in stored state. IAM is important, but IAM alone does not solve these problems.

Duplicate side effects

Durable execution does not make external actions automatically safe. If a workflow resumes after a timeout, it may repeat an email, payment, ticket, database write, or deployment.

Production designs should use:

• Idempotency keys
• Explicit transaction boundaries
• Approval checkpoints for consequential actions
• Compensating actions
• Retry budgets and dead-letter queues
• Side-effect inventories
• Human escalation paths

Regional limits

The cited GPT-5.6 availability is concentrated in US regions. Organizations with European, Asian, government, or regulated workloads must verify endpoint support, cross-region behavior, data-transfer paths, residency commitments, service tiers, and model-specific feature parity.

Runtime availability

GA availability of models does not prove GA availability of the separately announced Stateful Runtime Environment. Keep these claims separate in architecture reviews, budgets, and procurement documents.

Should a company choose Bedrock or OpenAI directly?

Bedrock is the stronger starting point when the company already operates heavily on AWS, needs IAM and VPC-oriented governance, wants CloudTrail and AWS-native operations, must connect agents to AWS-hosted data and services, or can use AWS commitments to simplify procurement.

Direct OpenAI access may be preferable when the team wants the newest OpenAI capability immediately, is not AWS-centric, needs the simplest direct API relationship, depends on OpenAI-specific features not yet exposed on Bedrock, or wants to minimize AWS-specific coupling.

Another Bedrock model may be better when cost, latency, language support, context behavior, regional availability, or tool performance favors Anthropic, Amazon Nova, or another provider. A multi-model fallback strategy can also reduce dependence on one model supplier.

1. Inventory the AWS services, private systems, identities, and commitments the agent must use.
2. Confirm the exact model, region, quotas, API features, and data-handling terms.
3. Price the complete workflow, including runtime, storage, retrieval, tools, logs, retries, and human review.
4. Design state export, deletion, retention, and recovery procedures before production.
5. Test duplicate side effects and approval boundaries with realistic failure scenarios.
6. Compare the same workflow on Bedrock and the direct OpenAI platform, measuring reliability, latency, cost, and operational burden.
7. Keep application-level state and tool contracts portable where strategic flexibility matters.

The broader implication

OpenAI and AWS are competing for different but complementary layers of enterprise AI. OpenAI retains the model technology, Codex, agent harness, and Frontier platform. AWS supplies the cloud distribution channel and can provide the identity, networking, infrastructure, audit, procurement, and governance environment in which agents operate.

That makes the partnership more consequential than a reseller arrangement. It is an attempt to make the cloud platform—not only the model provider—the place where enterprise agent execution is defined and governed.

But the headline needs a factual correction: OpenAI and AWS announced a stateful runtime, while the separately documented GA releases cover OpenAI models and Codex on Bedrock and, later, GPT-5.6 models. The full runtime’s production availability, regions, quotas, and standalone commercial terms should be verified before a company treats it as an established platform.

Quick Recap

Bestseller No. 1

Nimo AI NAS, Agentic Computer Mini PC and AI Server, AMD Ryzen 7 PRO 8845HS(up to 5.1 GHZ, beat i5-1235u) up to 132TB ZFS Hybrid Storage, Dual 10GbE for 24hr AI Agent

$1,999.99

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.