OpenAI DevDay 2026: Agents API Gets Computer Use, Codex Moves to the Cloud

Nov 10, 2026 · 6 min · Ajay Kumar

OpenAI held DevDay 2026 in San Francisco on 29 September and, by its own count, shipped more than 20 announcements in one keynote. For engineers the core of it was infrastructure: the Agents API now does computer use, multi-agent orchestration and context compaction on servers OpenAI runs, Codex can run in cloud environments rather than only on your laptop, and AWS launched Bedrock Managed Agents powered by OpenAI the same day. I build sandboxes for agents, so I read this as OpenAI deciding it wants to own the execution layer, not only the model.

What happened at OpenAI DevDay 2026

Axios focused its roundup on the headline products (always-on "dots" agents, ChatGPT Space and the Sol model). The developer platform changes, as reported by InfoQ and Reworked, were these:

Agents API: computer use (operating software through its GUI), multi-agent capabilities taken from Codex, tool search, tool calling and context compaction, with OpenAI managing the execution infrastructure.
Codex in the cloud: Codex tasks can run in cloud environments as well as locally, so you can start remote work from another device. The Codex CLI gained voice input and an /agents view for delegating and monitoring several tasks.
Code review and Codex Security Cloud: a review workflow for GitHub pull requests and GitLab merge requests, plus a service that scans repositories and new commits, deduplicates findings and prepares fixes.
Decisions API (limited preview): a constrained endpoint that picks from a predefined set of answers using the smaller Luna model, aimed at classification, routing and choosing an agent's next action.
ChatGPT plugins: sidebars, interactive panels and custom file viewers, with event handling through a proposed MCP Events specification.
Pro 500: a $500 a month ChatGPT plan with a usage allowance of 25x Plus and an "Ultrafast" tier that OpenAI says generates tokens up to 8x faster in Codex and up to 6x faster in the API.
Dots: persistent agents, in beta for Enterprise, Edu and Healthcare workspaces, off by default until an admin enables them.

The model news (GPT-6.1 Sol, priced at one fifth of GPT-6 Astra's standard token rates per InfoQ) belongs to a different story; what matters here is that Sol is pitched at coding and computer use, the two things the Agents API now hosts.

The numbers

Item	Detail	Source
Event date	29 September 2026, San Francisco	Axios
Pro 500 price	$500 a month, 25x the Plus allowance	Reworked
Ultrafast speed claim	up to 8x faster in Codex, up to 6x in the API	Reworked
GPT-6.1 Sol cached input	$0.10 per million tokens	InfoQ
Dots integrations	more than 4,000 apps via ChatGPT plugins	Reworked
Bedrock Managed Agents regions	us-east-1, us-west-2, us-east-2 (preview)	AWS docs
Bedrock Managed Agents price	no extra charge in preview beyond underlying AWS resources	AWS What's New

Bedrock Managed Agents: the AWS half of the launch

The AWS piece is the one I would read first. Bedrock Managed Agents (BMA) is built on a customised version of the Agents API, made AWS-native. According to the AWS documentation, the service holds the conversation and model calls, while tools execute in compute you provide: either your own host or Amazon Bedrock AgentCore Runtime. The bridge is a codex exec-server process that makes an outbound connection to the service. Each session assumes its own IAM role, supports MCP servers and reusable skills, can require human approval before consequential actions, and logs supported API activity to CloudTrail.

The preview uses a new endpoint family and signing service:

# Preview endpoint per AWS docs (us-east-1 shown). Requests are SigV4-signed
# with the service name "bedrock-mantle"; sessions live under this path.
BMA_ENDPOINT="https://bedrock-mantle.us-east-1.api.aws"
SESSIONS_PATH="/openai/v1/agents/sessions"
MODELS_PATH="/v1/models"   # model discovery; availability varies by Region and account

The AWS weekly roundup on 5 October also listed GPT-6.1 Sol and an Ultrafast mode for GPT-6 Astra as new models on Bedrock, so the models and the harness arrived together.

Why it matters

A year ago the usual pattern was: call a model API, run the agent loop yourself, and execute tools in a sandbox you picked. DevDay moved two of those three pieces to OpenAI. The Agents API now owns the loop and, with computer use, a managed machine to act on. Codex in the cloud does the same for coding work. That is simpler to start with and harder to inspect, which is the trade every managed service offers.

The AWS deal is the counterweight. BMA keeps the loop managed but puts tool execution, files and identity inside your own account. For teams that already gate everything through IAM and CloudTrail, that is a far easier security review than sending a hosted agent onto the internet with credentials.

Disclosure: I run PandaStack, an open-source microVM platform for agent code execution, so I have a stake in where tool execution lives. With that said, BMA's split (managed reasoning, customer-owned execution) is the architecture I would have recommended anyway.

What sceptics say

Most of the speed and capability claims are OpenAI's own. Reworked pointed out that the Ultrafast multipliers describe token generation, not end-to-end task time, and agent tasks are often dominated by tool calls and waiting. Several products are not generally available: the Decisions API is in limited preview, Dots is an opt-in beta for some plan types, and BMA is a public preview whose pricing AWS says "is subject to change at general availability." OpenAI's own recap page could not be fetched for this article, so the details above rely on InfoQ, Reworked, Axios and AWS.

The other risk is lock-in by convenience. Context compaction, tool search and multi-agent orchestration living server-side means the behaviour of your agent depends on code you cannot read or pin.

What it means for engineering and DevOps teams

Decide where tools run before you pick an API. If your policy says code and credentials stay in your VPC, BMA with self-hosted compute or AgentCore Runtime fits that; the hosted Agents API with computer use does not.
Watch the hidden infrastructure bill. AWS warns that its AgentCore example creates a NAT gateway and storage that keep charging when no agent turn is running. Tear down preview stacks.
Pin and evaluate. Server-side compaction and routing can change between releases. Keep an eval suite on the tasks you care about and run it when OpenAI ships updates.
Price Pro 500 against API usage. For heavy Codex users the flat plan may beat metered API calls; for automation, it will not, because plan allowances are for people, not pipelines.

FAQ

When was OpenAI DevDay 2026?

DevDay 2026 took place on 29 September 2026 in San Francisco. OpenAI described it as its biggest DevDay, with more than 20 announcements across ChatGPT, Codex, models and the API.

What is Bedrock Managed Agents powered by OpenAI?

It is an AWS service in public preview since 29 September 2026, built jointly with OpenAI on a customised Agents API. Agents run OpenAI models on Bedrock with their own IAM role, while tools execute on your own host or on AgentCore Runtime. It is available in three US Regions with no extra charge during the preview beyond the AWS resources used.

What is the OpenAI Decisions API?

It is a limited-preview API that chooses an answer from a fixed set of options based on text or image input, using a small model. OpenAI positions it for classification, request routing and picking an agent's next step.

Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.

I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.

Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.

Nov 10, 2026