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

OpenAI used its DevDay 2026 event to roll out a broad package of developer-facing updates. The headline items are a new model called GPT-6.1 Sol, computer-use support in the Agents API, cloud-hosted Codex environments, a Decisions API, and an expanded plugin system for ChatGPT 12. These could be read as unrelated launches. Together, though, they show a company moving away from selling raw model access and toward running the infrastructure that agents depend on.

What was announced

Both accounts of the event list the same core lineup 12. The more detailed report goes deeper on the agent and coding tools than on the model itself 2. That emphasis is telling.

GPT-6.1 Sol and the Decisions API get top billing, but neither report gives benchmarks, pricing, or a functional description of either product 12. Developers should treat those two as announced names with specifics still to come. Their capabilities have not been documented.

The Agents API grows up

The most consequential change is to the Agents API. It can now perform computer use, meaning applications can drive software through its graphical interface rather than only through structured APIs 2.

The API also takes on several capabilities from Codex:

• Multi-agent orchestration
• Tool search and tool calling
• Context compaction, for managing long-running sessions

OpenAI will also manage the execution infrastructure underneath 2. Access runs through the API, and through Codex and ChatGPT Work on selected plans 2.

This looks like an attempt to absorb the plumbing that agent builders have been assembling themselves. Teams have typically stitched together sandboxed execution, tool registries, and memory management on their own. If OpenAI hosts that layer, building gets easier. The trade-off is that developers become more tied to OpenAI's runtime. That raises familiar questions about portability, observability, and cost control, and the announcements do not address them.

Computer use deserves its own caution. Letting software operate a GUI makes it possible to automate legacy applications that lack APIs. It also widens the space for errors and misuse in ways that structured tool calls do not.

Codex moves to the cloud

Codex can now run in cloud environments as well as on local machines. Developers can start remote tasks from other devices 2. The Codex CLI also gained:

• Voice input
• An /agents interface for delegating and monitoring several tasks at once 2

Two additions push Codex further into the software lifecycle. The first is a code-review workflow that examines diffs and flags potential problems in GitHub pull requests and GitLab merge requests 2. The second is Codex Security Cloud, which can scan repositories and incoming commits, investigate findings, deduplicate them, and prepare fixes 2.

These features put Codex in direct competition with established code-review and application-security tooling. The deduplication and fix-preparation steps stand out. Security teams commonly complain about noisy scanners, and these steps are aimed at that problem. Whether Codex's findings prove more actionable than existing tools is something only real-world use will show.

ChatGPT plugins become a platform

ChatGPT's plugin system also expanded. Developers can now build:

• Sidebar experiences
• Interactive conversation panels
• Custom file viewers 2

Plugins can also react to events through a proposed MCP Events specification. Automations can then fire when something happens in a connected application 2.

The MCP Events piece may matter more than the interface additions. Event-driven triggers change plugins from tools a user calls during a chat into background automations that act on outside signals. It is notable that OpenAI is working through a proposed extension to the Model Context Protocol instead of a proprietary hook. It suggests OpenAI sees value in interoperability at the integration layer, even as it centralizes execution elsewhere. As a proposal, though, the specification's final shape and level of adoption remain open.

Reading the announcements

The available reporting is consistent across outlets, since one account summarizes the other 12. It is also thin on model details and heavy on tooling. That imbalance seems to reflect where OpenAI wants developers looking.

The common thread is consolidation. Execution environments, multi-agent coordination, code review, security scanning, and event-driven automation are all moving inside OpenAI's managed services.

For developers, the practical question is less about the new model and more about how much of their agent stack to hand over. The convenience is real: hosted execution, cloud Codex sessions, and richer plugin surfaces remove a lot of integration work. The cost is dependence on one vendor for compute, orchestration, and distribution. That bargain will look very different to a startup moving quickly than to an enterprise with compliance and portability requirements.

Until OpenAI publishes concrete details on GPT-6.1 Sol, the Decisions API, and pricing for managed execution, the most defensible takeaway is this: DevDay 2026 was mainly about OpenAI becoming the place where agents run, not just the provider of the models behind them.