Two announcements in the last week rearranged the AI options for anyone building on AWS, which is most of our clients.
On Monday 27 April, OpenAI and Microsoft rewrote their agreement. Microsoft's licence to OpenAI's technology is now non-exclusive, and OpenAI can serve all of its products on any cloud. The next day AWS said the obvious out loud: OpenAI's latest models, its Codex coding agent and a new managed agent service are coming to Amazon Bedrock. Then on Thursday AWS quietly announced the end of Amazon Q Developer, its own coding assistant, and pointed everyone at Kiro.
Here is what each of those means for a startup on AWS, and what we are changing in the architecture overviews we hand clients at the end of a Flight Plan.
What Actually Landed
Three things on Bedrock, all in limited preview:
- OpenAI models. GPT-5.5 and GPT-5.4 through the Bedrock API, with what AWS calls unified security, governance and cost controls. In plain terms, the same IAM permissions, the same bill and the same logging as every other model you already use there.
- Codex. OpenAI's coding agent, starting with the command-line tool, the desktop app and a VS Code extension, with inference routed through your Bedrock account.
- Bedrock Managed Agents powered by OpenAI. A hosted way to build agents on OpenAI models with AWS infrastructure underneath.
Amazon's CEO said the models would be available directly to customers on Bedrock "in the coming weeks". The rest of the deal explains why it happened now: Amazon has committed up to $50 billion to OpenAI, the two will co-develop a stateful runtime on Bedrock for long-running agents, and AWS gets exclusive rights to OpenAI's Frontier agent-building product. Microsoft keeps a first-on-Azure clause for new OpenAI releases and a 27 percent stake, so expect a lag between an Azure launch and a Bedrock one.
And the retirement: new Amazon Q Developer subscriptions stop on 15 May 2026, Claude Opus 4.6 leaves Q Developer Pro on 29 May with newer models going to Kiro only, and full end of support is 30 April 2027. The IDE plugins stop working then. Q Developer inside the AWS console carries on. The replacement is Kiro, AWS's spec-driven development environment.
Why This Matters for a Serverless Client
We build client backends on Lambda, API Gateway, DynamoDB and S3, and where a product needs a language model we have reached for Bedrock whenever we could, because it keeps the model call inside the same account, region and permission boundary as everything else. The gap was that Bedrock had Anthropic, Amazon, Meta, Mistral and a long tail, but not OpenAI. A client who wanted GPT for one feature needed a second vendor, a second API key in a secrets manager, a second invoice and a second data-processing agreement.
That gap is closing. Once the preview opens up, an AWS client can put a Claude model and a GPT model behind the same interface and choose per feature. For the AI Automation work we do, that is useful: a classification step on one model, a drafting step on another, with the choice made by an eval rather than by which vendor signed the contract.
There is a second, quieter benefit. A model swap becomes a configuration change inside the account, with no new procurement. We wrote about that kind of portability in March, when a provider dispute made the point for us. This week made it cheaper.
Codex Through Your Own AWS Account
The Codex part is the one we are most curious about. Until now, using OpenAI's agent on a client build meant an OpenAI account and OpenAI's billing, which some enterprise clients could not sign off on. Codex on Bedrock means the same CLI and desktop app, billed per token through the client's existing AWS account and covered by whatever data agreements they already have with AWS.
For a Takeoff & Climb build inside a client's AWS organisation, that removes the awkward vendor conversation entirely. It also gives us a like-for-like comparison: Claude Code and Codex, same repository, same account, same bill, and the eval decides which does what.
The Q Developer to Kiro Move
If you already have a product on AWS and your developers used Q Developer in their editor, put a migration on the Cruise & Maintenance list now. There is a year, but the newest models leave it within the month, so it will get worse before it stops.
Kiro is a reasonable landing. We have watched it since its autonomous agent preview at re:Invent, and its whole design is spec first: a requirements document, a design document and a task list before the agent writes code. That is the process we already run, so for a client's own developers the transition is mostly learning a new tool, not a new way of working.
What We Are Not Assuming Yet
Pricing. AWS has not said what OpenAI models cost on Bedrock. Until it does, we quote OpenAI usage at OpenAI's own rates and note the Bedrock option as pending.
Regions. Nothing has been said about Sydney. Bedrock model availability in Australia has lagged the US for every launch we can remember. For clients with residency requirements, this stays on the "when" list.
Timing. "Limited preview" and "coming weeks" are AWS phrases with some stretch in them. We are not planning a build around a service that is not generally available.
What Changes in Our Flight Plans
Starting this week, the architecture overview for an AWS client names Bedrock as the model layer and lists both Anthropic and OpenAI models behind it, with the caveats above. Where a client already has an OpenAI relationship, we note the option to consolidate. Where a client uses Q Developer, the roadmap gets a Kiro migration item.
One account. One bill. Both frontier labs. For a startup trying to keep its operations simple, that is the week's real headline.
Building on AWS and Weighing Your Model Options?
At Dev Force One, we design serverless backends on AWS and put the model layer inside the same account as everything else, so the choice of model stays a configuration decision rather than a procurement one.
What we bring:
- Bedrock-first architecture with Anthropic and now OpenAI models behind one interface
- Evals, not opinions: model choices per feature decided by tests on your data
- A migration path from Amazon Q Developer to Kiro for existing products
Planning a product on AWS, or rethinking the model layer of one you already run? Book a Pre-flight call and we will map it with you.