OpenAI Lands on AWS: What Bedrock Managed Agents Mean for Businesses Building AI Agents in 2026

On April 28, 2026 AWS switched on three things at once, all limited preview: OpenAI's latest models on Bedrock, Codex on Bedrock, and a new product called Bedrock Managed Agents built on the OpenAI agent harness. It only happened because OpenAI got Microsoft to drop its API exclusivity clause the day before, in exchange for a non exclusive license through 2032 plus a revenue share.

Update, October 10, 2026: Two of the three products are now generally available. GPT-5.5, GPT-5.4 and Codex on Bedrock went GA on June 1, 2026, and the catalog has since added the GPT-5.6 family (July 13) and the GPT-6 family, including GPT-6 Astra (September 8) and GPT-6.1 Sol (September 29). AWS says pricing matches OpenAI's first party rates and counts toward existing AWS commitments. Bedrock Managed Agents moved from limited preview to a public preview on September 29, 2026, in the N. Virginia, Ohio and Oregon Regions, with no extra charge beyond the AWS resources it consumes. AWS has not announced a GA date. One correction: OpenAI's own April 27 statement says it can now serve all its products to customers on any cloud, with products shipping first on Azure. This post originally described Frontier as Azure exclusive through 2032, and the published terms do not say that. What is still true is that Frontier is not sold through AWS. Every affected section below is updated in place.

The headline item is Managed Agents: an opinionated agent runtime with identity per agent, audit logging, and AgentCore underneath, with Box as the named launch customer. For anyone building agents this changes procurement more than architecture, because you can now buy OpenAI reasoning models, Codex, and a managed agent platform without leaving your AWS account or your existing spend commitment. The catch worth knowing before you plan around it: Frontier, OpenAI's own first party agent product, still lives on Azure. What landed on AWS is OpenAI's frontier models behind AWS controls, not Frontier the product.

What actually launched on April 28, 2026?

Three products, all stamped "limited preview," all announced in a single Amazon staff post on aboutamazon.com. The headline you're hearing, "OpenAI is on AWS now," is correct, but the shape of the launch is more specific than the framing suggests.

Here's the literal list, pulled from Amazon's announcement:

OpenAI models on Amazon Bedrock. The latest OpenAI frontier models, accessible through the same Bedrock APIs and IAM controls AWS customers already use. They show up next to Anthropic, Meta, Mistral, Cohere, and Amazon's own Nova models in Bedrock's model catalog. One unified billing surface, one set of guardrails, one cost-control layer. If you already have an AWS Enterprise Discount Program or a Marketplace private offer, OpenAI usage drops into that envelope instead of becoming a separate vendor relationship.

Codex on Amazon Bedrock. OpenAI's coding agent now runs in your AWS environment. Authentication uses your AWS credentials. Inference goes through Bedrock infrastructure. And here's the line that matters for procurement: "Customers can authenticate using their AWS credentials, process inference through Amazon Bedrock infrastructure, and apply Codex usage toward their AWS cloud commitments." Translation: if your finance team already negotiated a 3-year AWS commit with a discount, Codex now burns those dollars instead of opening a new ChatGPT Enterprise contract. At launch it was available through the Bedrock API, the Codex CLI, the Codex desktop app, and a VS Code extension. Since the June 1, 2026 GA, AWS lists the Codex App, the Codex CLI, and IDE integrations for Visual Studio Code, JetBrains and Xcode, billed per token with no seat licenses.

Amazon Bedrock Managed Agents, powered by OpenAI. This is the new product, not a port. It's a managed agent runtime designed specifically for OpenAI's reasoning models. Amazon's launch copy promised agent steering, security, identity per agent, memory that persists across sessions, and a log of every action. The September 2026 preview delivers durable sessions, one IAM role per agent and CloudTrail records of supported API activity, but lists no built in long term memory integration yet. The first big customer Amazon named is Box, which has 115,000 organizations on its content management platform.

At launch OpenAI said more than 4 million people used Codex every week, and by its June 1 GA announcement that was more than 5 million. Putting that user base inside enterprise AWS accounts is a serious wedge.

Where it stands in October 2026: models, Regions and pricing

The models and Codex left limited preview after 34 days. Here is the OpenAI lineup on Bedrock as of October 10, 2026, taken from the AWS What's New posts, the Bedrock OpenAI model cards and the Bedrock Regions page, which lists where each model is hosted in Region.

The GPT-5.6 and GPT-6 models also support Geo and Global cross Region inference, added in August 2026, which lets you call them from many more commercial Regions while Bedrock routes each request to a Region where the model runs. GPT-5.5 and GPT-5.4 are in Region only. So if a regulator wants strict single Region processing outside N. Virginia, Ohio or Oregon, the proprietary OpenAI models are not there yet. The open weight gpt-oss models are the exception, hosted in Region across Europe, Asia Pacific and South America.

On price, AWS's position is simple: you pay the same per token rate as buying direct from OpenAI, with no additional fees, and the spend counts toward your existing AWS commitment. Codex on Bedrock is pay per token with no seat licenses. GPT-5.6 supports explicit prompt caching, and AWS says repeated context on it is billed at a 90% discount. The Bedrock Regions page also notes that for some models, Global cross Region inference is priced lower per token than Geo routing. Current per token rates are on the Bedrock pricing page.

Why did this take so long? The Microsoft exclusivity story

The Bedrock launch is the easy part. The deal that made it possible is the interesting part.

In November 2025, OpenAI signed a $38 billion multi-year cloud contract with AWS. In February 2026, Amazon announced an investment of up to $50 billion in OpenAI ($15 billion upfront, $35 billion contingent on undisclosed conditions), in exchange for which OpenAI agreed to two things: jointly develop "stateful runtime technology" on AWS Bedrock (the infrastructure that lets agents remember tasks across long running sessions), and make AWS the exclusive third party cloud distribution provider for OpenAI's new agent product, Frontier.

That second part broke the deal. Microsoft's existing agreement with OpenAI gave it exclusive rights to any OpenAI product accessed through an API. The day Amazon announced its investment, Microsoft put out a public statement refuting the AWS exclusivity, in bold: "Azure remains the exclusive cloud provider of stateless OpenAI APIs. Any stateless API calls to OpenAI models that result from a collaboration between OpenAI and any third party, including Amazon, would be hosted on Azure. OpenAI's first party products, including Frontier, will continue to be hosted on Azure."

So OpenAI had taken Amazon's $50 billion to do something Microsoft was, on paper, blocking it from doing. That's the legal peril TechCrunch's headline refers to.

On April 27, 2026, that knot got cut. Under the amended agreement, Microsoft keeps its license to OpenAI IP for models and products through 2032, but the license is now non exclusive. Microsoft stops paying a revenue share to OpenAI, while OpenAI's revenue share payments to Microsoft continue through 2030 at the same percentage, subject to a total cap. In return, OpenAI got the right to actually sell its products on AWS: in OpenAI's words, it "can now serve all its products to customers across any cloud provider," with products shipping first on Azure. Microsoft is still being called OpenAI's "primary cloud partner" in both companies' statements, which is likely to stay true in volume terms (Azure should keep serving the bulk of OpenAI's API load), but the exclusivity is gone.

One day after Microsoft signed off, Amazon turned on three products. CEO Andy Jassy posted a short note on X about the Microsoft news ("very interesting announcement") on Monday, then on Tuesday Amazon's blog quietly went live. That timing is not a coincidence. AWS clearly had this build ready to ship.

What is Bedrock Managed Agents, really?

Of the three launches, this is the one worth the most reading time. The other two are competitive rebalancing. This one is a new product category for AWS.

Bedrock Managed Agents is built on the OpenAI agent harness, which is OpenAI's runtime for orchestrating long-running, multi-step tasks. Amazon describes it as "engineered to unlock the full potential of OpenAI frontier models, delivering faster execution, sharper reasoning, and reliable steering of long-running tasks." Translation: it knows how to drive OpenAI's reasoning models specifically, with the kind of mid-task correction that generic agent runtimes fumble.

Three operational details matter for anyone evaluating this for production:

Identity per agent. Every agent gets its own AWS identity. That means least-privilege IAM policies attached directly to the agent, not to the human who launched it. If an agent is compromised, you blast-radius it the same way you'd blast-radius any service principal. This is a real difference from running OpenAI's Responses API in your own infra (the older Assistants API was shut down on August 26, 2026), where identity bookkeeping is your problem.

Audit logs. The launch marketing copy said it "logs every action for auditability." The September 29, 2026 preview announcement is narrower: each agent records supported API activity with AWS CloudTrail. The actual value is that the record lands in CloudTrail, which you can already route to S3 or any SIEM you have, so the API layer of an audit needs no custom telemetry. Tools run in the compute environment you provide, so logging what they do there is still your job.

AgentCore as the compute layer. Bedrock Managed Agents runs on top of Bedrock AgentCore, AWS's open agent platform. Amazon says, "AgentCore provides the default compute environment for Bedrock Managed Agents, and as your agent footprint expands across your enterprise, AgentCore and Bedrock Managed Agents will provide additional capabilities such as authorization policy enforcement, agent and agent tool discovery, and observability and evaluation capabilities." The interesting word there is "expands." Amazon is saying: build with Managed Agents now, scale into AgentCore for enterprise discovery and governance, and never have to switch runtimes. In the September preview, AgentCore Runtime is one of two execution environments. The other is self hosted compute: an existing machine, container or compute environment that runs the BMA exec server.

Box's CTO Ben Kus called it a way to "build optimized, production-scale AI applications" without the integration tax. That's PR copy, but Box's base of more than 115,000 customer organizations is real.

What the September 2026 public preview actually ships

Five months after the April announcement, AWS opened Bedrock Managed Agents (BMA) as a public preview on September 29, 2026. AWS now describes it as a customized version of OpenAI's Agents API engineered to be AWS native, and the product page says it combines OpenAI models with the Codex harness and AgentCore. Here is what you get, per the Bedrock documentation, checked October 10, 2026:

Included: durable sessions that keep messages, tool calls and intermediate results so an agent can resume later; reusable skills; tools through MCP servers; one IAM role per agent; human approval before consequential actions; CloudTrail records of supported API activity; and a choice of self hosted compute or AgentCore Runtime for tool execution.

Not included yet: the preview limits page lists OpenAI models only, no subagents, no built in long term memory integration, text input only, API access with no console procedures, no cross Region inference profiles, and no customer managed KMS key setting.

Where and what it costs: N. Virginia, Ohio and Oregon, through the bedrock-mantle endpoint. During the preview there is no additional charge for BMA beyond the AWS resources your agents consume, and AWS says pricing is subject to change at general availability.

AWS Bedrock vs Azure for OpenAI: where does the choice actually matter?

Here's the comparison most coverage missed. Both clouds now serve OpenAI models. They're not interchangeable.

If you're already deeply on Azure with a 3-year ECP, this announcement changes very little for you. Azure stays OpenAI's primary cloud and the first place new OpenAI products ship, and Frontier is still not sold through AWS. If you're on AWS with a multi-year EDP and you've been waiting to consolidate AI vendors, the consolidation is now possible, and that's the actual news. For everyone in the middle, you finally have real model choice. You can put GPT-class reasoning, Claude, Llama, Mistral Large, and Nova into the same Bedrock-managed app and route by task. That used to require separate API integrations.

If you're trying to figure out where to actually start, our guide to AI agent development services walks through the procurement-vs-build decision in more detail.

What this means for businesses building AI agents

I've now built or shipped 126 production AI systems. The most common reason agent projects stall isn't the model. It's the integration tax. You pick OpenAI for the model quality, Pinecone or Weaviate for vectors, LangChain or LlamaIndex for orchestration, Datadog for observability, Auth0 for identity, and somewhere in there your security team asks for an audit log and you realize you have to build one.

Bedrock Managed Agents is Amazon's bet that this integration tax is worth eliminating, even at the cost of some flexibility. For most enterprise teams, that's the right tradeoff. The same logic that pushed companies onto Bedrock AgentCore in late 2025 applies more strongly here: a managed runtime with built-in identity, observability, and policy enforcement saves quarters of work compared to assembling the equivalent stack yourself.

Three things change for anyone shipping agents in the second half of 2026:

Procurement gets simpler. If your company has an AWS EDP, you stop opening new vendor relationships for OpenAI usage and Codex (which bills per token on Bedrock, not per seat). That's not a minor thing. Vendor management overhead is real, and security review for a brand-new SaaS contract can take 60-90 days at large enterprises. Buying OpenAI inside your existing AWS contract collapses that into an IAM permission change.

Multi-model agents get easier. Bedrock has been pushing the "model garden" pitch for two years. Now it has OpenAI, which was the biggest gap. Google's Gemini is still missing as of October 2026: the Google models on Bedrock are the open weight Gemma family. You can A/B test Claude vs GPT vs Llama on the same task, with the same observability, without three integrations. If you've been wrestling with whether to use Claude for reasoning and GPT for coding, the answer is now "both, and route at the orchestration layer." For builders comparing the alternatives, our n8n vs Zapier breakdown covers the no-code side of that same routing problem.

Compliance posture improves. The "agent has its own IAM identity" model is what enterprise security teams actually want. Today, most production agents share a service account or a human's keys, which is unauditable. Bedrock Managed Agents flips that. If you've been blocked from production by InfoSec, this is the first agent runtime they'll like.

If you're newer to all of this, our honest 2026 guide to agentic AI covers the basics without the hype.

Where the announcements fall short (what TechCrunch and Amazon both missed)

Two things nobody is saying loudly.

First, "limited preview" was doing a lot of work. At launch on April 28, 2026, all three products sat behind a sign up form with no public pricing, no published rate limits and no SLA. Two of them cleared that bar quickly: the models and Codex went GA on June 1, 2026, 34 days later. Managed Agents did not. It reached a public preview on September 29, 2026, still with no GA date and with pricing "subject to change at general availability." For scale, AgentCore itself went from preview on July 16, 2025 to GA on October 13, 2025, about three months, and Managed Agents has now been short of GA for more than five. If your roadmap depends on a launch date, do not bake Managed Agents into the critical path until you have a date in writing.

Second, Frontier is still on Azure. This is the part of OpenAI's product line that's furthest into the agent era. It's the new agent-making tool that the entire $50B Amazon deal originally hinged on. As of October 2026 it is still not among the OpenAI products sold through AWS, even though AWS is its named exclusive third party cloud distributor and OpenAI's April 27 terms let it serve all its products on any cloud. So when the marketing says "OpenAI on AWS," the asterisk is: OpenAI's models, OpenAI's coding agent, and OpenAI's agent harness inside an AWS-built runtime. Not OpenAI's flagship agent product itself. If you specifically want Frontier, you still go through OpenAI, not AWS.

Andy Jassy's "very interesting announcement" tweet was, I think, deliberately understated. AWS got the parts of OpenAI it could productize within AWS's existing platform model. It didn't get the part that competes with AWS's own agent platform. That's not a coincidence either.

For broader agent-platform context, our recent A2A protocol coverage covers the cross-vendor agent communication standard that this announcement now plugs into more directly.

Should you wait for Bedrock Managed Agents or ship now?

Short answer: if you're starting an agent project now, prototype on whatever stack you can move fastest with. Options include LangGraph, CrewAI, raw OpenAI Agents SDK, or Anthropic's Claude Agent SDK on Bedrock. Get to a working agent. Then, when Managed Agents hits GA (AWS had not announced a date as of October 10, 2026), evaluate migrating only if you're hitting governance or scale problems your current stack can't solve.

The reasons not to wait:

Preview means feature changes. APIs in preview can and do break before GA. Building a production system on a preview runtime is a known anti-pattern. AWS says so itself on the Managed Agents docs page: "Preview functionality and APIs can change."

The integration tax is real but it's not infinite. A senior backend engineer can wire OpenAI Agents SDK to Postgres, S3, and CloudWatch in two weeks. You don't need a managed agent runtime to get to production. You need one to scale from 5 agents to 500.

The reasons to wait, or at least to plan around it:

If your security team won't approve a new SaaS vendor, Bedrock Managed Agents at GA may be the only viable path. Your timeline is dictated by their release schedule.

If you're building 50+ agents across the org, the AgentCore plus Managed Agents combo is genuinely better than rolling your own. Don't fight that battle.

If you're already building on AgentCore, Managed Agents can use AgentCore Runtime as its execution environment, so the switching cost is low, as long as you are happy running OpenAI models only.

Personally, on the projects we're shipping in May 2026, we're prototyping on raw OpenAI Agents SDK and Bedrock AgentCore in parallel, with a portability layer that lets us move whichever way the GA dates fall. That's not a sexy answer, but it's the one that doesn't blow up a roadmap.

Frequently asked questions

Is OpenAI on AWS Bedrock generally available?

Mostly. OpenAI models and Codex on Bedrock have been generally available since June 1, 2026, starting with GPT-5.5 and GPT-5.4. The GPT-5.6 family followed on July 13 and the GPT-6 family in September 2026, with GPT-6.1 Sol GA on September 29. Bedrock Managed Agents, powered by OpenAI, is the exception: it moved from limited preview to public preview on September 29, 2026, and as of October 10, 2026 AWS has not announced a GA date.

What is Bedrock Managed Agents, powered by OpenAI?

It's a new managed runtime for building production AI agents using OpenAI's frontier models. It runs on AWS infrastructure, is built on a customized version of OpenAI's Agents API, gives each agent its own IAM role, records supported API activity in CloudTrail, and runs tools on AgentCore Runtime or on compute you host yourself. It is in public preview in N. Virginia, Ohio and Oregon, with no extra charge beyond the AWS resources it uses. Box, with 115,000 customers, was the named launch reference.

Does this mean Microsoft Azure loses OpenAI?

No. Microsoft keeps a license to OpenAI IP for models and products through 2032, now non exclusive, and OpenAI keeps paying Microsoft a revenue share through 2030, subject to a total cap. Microsoft is still called OpenAI's "primary cloud partner" by both companies, and OpenAI products ship first on Azure. The change is that OpenAI can now serve its products on any cloud, AWS included. Frontier (OpenAI's first party agent product) is not sold through AWS as of October 2026, and Microsoft said in February that it stays hosted on Azure.

Can I use Codex on Bedrock as a drop-in for ChatGPT Enterprise?

No. Codex on Bedrock is the OpenAI coding agent, not the full ChatGPT Enterprise product. Since June 1, 2026 it has been generally available through the Codex App, the Codex CLI, and IDE integrations for Visual Studio Code, JetBrains and Xcode, billed per token with no seat licenses. Inference goes through your AWS account and counts toward your AWS commit. Separately, AWS says ChatGPT Work can be configured to use GPT-6 Astra on Bedrock.

How does AWS Bedrock Managed Agents compare to Azure AI Agent Service?

Microsoft now calls it Foundry Agent Service. It is generally available, works with many models from the Foundry catalog (Microsoft lists GPT-4o, Llama and DeepSeek as examples), and uses Microsoft Entra identity and RBAC. Bedrock Managed Agents is in public preview, runs OpenAI models only, is built on a customized version of OpenAI's Agents API, has tighter integration with AWS native security (one IAM role per agent, CloudTrail records), and runs tools on AgentCore Runtime or your own compute. If you're an AWS shop that wants OpenAI models, Bedrock wins on operational fit. If you need a GA managed agent service today, or several model vendors inside one agent runtime, Foundry Agent Service or Bedrock AgentCore are the GA options.

What is OpenAI's $50B Amazon deal in plain terms?

In February 2026, Amazon committed up to $50 billion in OpenAI ($15B initial, $35B contingent on conditions). In exchange, OpenAI agreed to jointly develop stateful runtime tech on AWS Bedrock and to make AWS the exclusive third party cloud distribution provider for OpenAI's Frontier agent product. The Frontier exclusivity broke against Microsoft's existing rights, which forced the April 27 renegotiation. The investment money is separate from OpenAI's prior $38B AWS cloud spend commitment from November 2025, which the same February deal expanded by $100 billion over eight years.

Should I move my OpenAI workload from Azure to AWS today?

For plain model calls, it is now a reasonable move if you're an AWS first shop: the models and Codex are GA on Bedrock, priced the same per token as OpenAI direct, and count toward your AWS commit. For agents, probably not yet. Bedrock Managed Agents is in public preview with no SLA or GA date, while Azure remains GA, mature, and, per Microsoft, the host for Frontier. The right move is to pilot Managed Agents on Bedrock while keeping production agent runtimes where they are until it hits GA.

The procurement story is the real story

If you only remember one thing from this announcement: the technology layer didn't move much. The procurement layer did.

OpenAI's models were already great. Codex was already great. The OpenAI agent harness already existed. What changed on April 28, 2026 is that AWS customers can buy all of it through the contract they already have, with the IAM model they already use, applied to the cloud commit they already negotiated, audited by the SIEM their security team already approved.

For a CTO at a 5,000-person company on AWS, that's the unblock. For a startup founder spinning up an agent demo this weekend, it's a footnote. Read the framing accordingly.

If you want a starting point on what to actually build first, take 5 minutes with our agentic vs generative AI decision guide or the broader 2026 agentic AI cheat sheet. Both are written from the same set of 126 production builds, and both will save you a planning meeting.

Three primary sources for this analysis. AWS & OpenAI announce expanded partnership (Apr 28, 2026) · TechCrunch: Amazon is already offering new OpenAI products on AWS (Apr 28, 2026) · TechCrunch: OpenAI ends Microsoft legal peril over its $50B Amazon deal (Apr 27, 2026) · Microsoft & OpenAI joint partnership statement (Feb 27, 2026).

Built 126 AI systems for businesses across SaaS, healthcare, professional services, and ops. If you want a no-pitch read on whether your business should be using OpenAI on AWS, Anthropic, or open-source models, take the 2-minute AI Readiness Quiz and we'll send a tailored recommendation.