OpenAI held DevDay 2026 in San Francisco on September 29, with announcements spanning models, agents, decision systems, developer infrastructure, privacy, identity and workplace collaboration.
There was no single announcement that defined the event.
Instead, the bigger message was that OpenAI is expanding beyond providing models. It is increasingly providing the infrastructure developers need to build, run and distribute AI applications and agents.
Here are the announcements that matter most.
1. GPT-6.1 Sol brings more capability to a lower-cost model
OpenAI introduced GPT-6.1 Sol, positioned for complex coding, computer use and professional work at a lower cost than GPT-6 Astra.
The model has a 1.05-million-token context window and supports up to 128,000 output tokens. Standard API pricing starts at $2 per million input tokens and $10 per million output tokens for supported prompt sizes.
The important development is not simply another model release.
OpenAI is continuing the pattern of moving capabilities closer to its frontier models into lower-cost tiers.
Why it matters: capable AI is becoming less expensive to use in real applications, especially where developers need large context, coding or agentic workloads.
This also reinforces a broader trend we have discussed separately on TechiesJournal: access to strong AI intelligence is becoming less scarce.
2. Ultrafast makes latency a product choice
OpenAI also introduced Ultrafast, a premium inference tier.
OpenAI reports up to 8× faster token generation in Codex and up to 6× faster generation through the API, with Astra available now and Sol support planned.
For many AI applications, model quality is only part of the experience.
A capable model that takes too long to respond may still be unsuitable for interactive coding, voice, real-time interfaces or agent workflows.
Why it matters: developers are increasingly choosing models not only by intelligence and cost, but also by latency.
3. Agents API moves long-running agents into managed infrastructure
The Agents API is now in public beta.
OpenAI manages the agent harness, context, durable sessions and orchestration while developers provide tools and decide where code should execute.
The platform can keep agents working across long sessions, manage context, coordinate subagents and use hosted or external execution environments. Computer use was also added, allowing agents to operate software through graphical interfaces.
This is one of the more important developer announcements.
Building an agent is relatively easy. Keeping one running reliably for hours or days, managing its state and recovering from interruptions is much harder.
Why it matters: some of the operational complexity of long-running agents is moving from individual application developers to managed AI platforms.
We will treat hosted agent infrastructure as a broader topic rather than repeat our recent Building Agentic Systems series.
4. Decisions API introduces a different kind of AI task
OpenAI announced the Decisions API in limited preview.
Powered by GPT-6 Luna, developers define questions and possible answers, and the system can classify information, route requests or choose an agent’s next action.
This is different from asking a language model to generate a long answer.
For many software tasks, the application already knows the possible choices:
route to team A or B
use tool X or Y
approve, reject or escalate
The model’s job is to make the decision.
Why it matters: AI development may be starting to separate generative intelligence from decision intelligence.
This deserves separate analysis because OpenAI is not alone. Jev and AWS Strands Decider are exploring the same broader problem from different technical directions.
5. Sign in with ChatGPT changes how applications can consume AI
Sign in with ChatGPT lets users authenticate to participating third-party applications with their ChatGPT identity.
More importantly, eligible Plus and Pro users can choose to allow supported applications to use AI requests against the usage included with their ChatGPT plan, without creating or sharing an API key. Users can also set usage limits for individual applications.
Identity sign-in itself shares only basic profile information such as name, email and profile picture. Conversations and ChatGPT memory are not automatically shared. Additional permissions require separate authorization.
Why it matters: this could change an important part of AI application economics.
Instead of every developer paying for all AI usage and recovering the cost through subscriptions, some applications may allow users to bring their existing AI entitlement with them.
This is one of the DevDay announcements we plan to examine separately.
6. OpenAI agents are coming to AWS infrastructure
OpenAI and AWS are also expanding their partnership through Amazon Bedrock Managed Agents powered by OpenAI.
The OpenAI agent harness can run through Amazon Bedrock, while AWS infrastructure handles execution, authentication and related enterprise controls.
OpenAI’s own comparison shows an important difference: with its native Agents API, OpenAI manages the agent loop and model inference. With Bedrock Managed Agents, the harness and model inference operate through Amazon Bedrock and AWS authentication.
Why it matters: enterprises may increasingly be able to use OpenAI capabilities without moving their broader infrastructure, governance and procurement model outside AWS.
It is another sign that AI providers and cloud platforms are becoming more interconnected rather than remaining completely separate stacks.
7. Private Intelligence focuses on the privacy-versus-safety problem
OpenAI grouped several privacy developments under Private Intelligence.
One important component is Zero Data Retention with Private Safety Processing.
The goal is to allow automated safety systems to detect risky patterns across interactions without giving OpenAI personnel access to the underlying customer content.
OpenAI says protected content can remain in customer-controlled storage, while only bounded safety signals leave the protected processing environment.
OpenAI also announced plans for a Private Inference preview.
Why it matters: enterprises handling sensitive information have long faced tension between data privacy and provider-level AI safety requirements.
Private Intelligence is OpenAI’s attempt to make those two requirements coexist.
This deserves deeper technical investigation once the Private Inference architecture is better documented.
8. Dots push ChatGPT toward persistent digital workers
OpenAI also introduced Dots, persistent agents with connected applications and their own cloud computer.
For enterprise environments, administrators can control access to Dots, connected apps, computer capabilities and messaging. Specialist Dots are also being tested for agents assigned specific organizational responsibilities.
This moves beyond asking ChatGPT a question.
The idea is closer to assigning ongoing responsibilities to an AI worker that remains available and connected to tools.
Why it matters: persistent agents are increasingly becoming products rather than demonstrations.
That raises larger questions around identity, governance, audit and control, exactly the infrastructure problem we recently examined through OpenClaw Enterprise.
9. ChatGPT is becoming a shared work environment
DevDay also expanded ChatGPT’s collaboration layer.
Announcements included:
- ChatGPT Space for shared files and project context
- Pages for collaborative documents
- Teams and Team Tasks
- Slack and Microsoft Teams integration
- Meetings for notes and action items
- collaborative slides
- shared recurring work
These features are less technically dramatic than a new API, but strategically they matter.
ChatGPT is moving from an individual assistant toward a place where humans and agents can share context, documents and recurring responsibilities.
10. Codex can keep working after you close your laptop
OpenAI also updated Codex Cloud.
Developers can create reusable cloud environments containing repositories, dependencies, scripts and settings. Codex tasks can continue running after the developer’s local computer is turned off and can be monitored or steered from another device.
Why it matters: coding agents are becoming persistent cloud workers rather than tools tied to an active developer session.
This is another part of the wider shift toward long-running AI work.
Which announcements matter most?
For developers, I would place the announcements into three groups.
Most important now
Agents API
Long-running agent infrastructure is becoming a managed platform capability.
Sign in with ChatGPT
Could change onboarding and the economics of third-party AI applications.
Decisions API
Points toward a potentially important new category of decision-oriented AI.
GPT-6.1 Sol
Continues the rapid decline in the cost of high-capability AI.
Important to watch
Private Intelligence
Potentially important enterprise architecture, but some components are still emerging.
Dots
Shows where persistent AI workers may be heading, particularly in enterprise environments.
OpenAI on AWS
Important for organizations that want OpenAI capabilities within existing AWS governance.
Useful platform improvements
Ultrafast, Codex Cloud, Teams, Pages, Meetings and collaboration features
These matter to users and developers, but they represent evolution of existing products more than entirely new architectural categories.
The broader DevDay message
The most interesting part of DevDay 2026 may not be any individual model.
OpenAI’s developer platform is expanding across several layers:
Accessible text alternative for this figure
A numbered stack of seven layers, read from top to bottom. 1. Models: GPT-6.1 Sol and Astra. 2. Decisions: Decisions API. 3. Agents: Agents API, computer use and Dots. 4. Execution: hosted sandboxes and Codex Cloud. 5. Identity and usage: Sign in with ChatGPT. 6. Enterprise infrastructure: AWS integration and Private Intelligence. 7. Collaboration: Spaces, Pages, Teams and connected applications. Each layer has a number and a text label, so no meaning depends on colour.
That is a much broader platform than a model API.
For developers, the question is increasingly shifting from:
Which OpenAI model should I call?
to:
Which parts of the application stack should the AI platform itself manage?
That is the larger development worth watching after DevDay 2026.
Topics We Will Explore Separately
Several announcements deserve more than a short event summary:
Sign in with ChatGPT
What changes when users can bring their existing AI subscription into another application?
Decision Models
How do OpenAI Decisions API, Jev and AWS Strands Decider differ, and does software need a generative model for every intelligent decision?
Private Intelligence
Can cloud AI preserve enterprise privacy while still supporting provider-level safety controls?
Other DevDay developments will remain part of our ongoing coverage rather than receiving separate articles unless the technology develops further.
Sources
Source review: 2 October 2026.
- OpenAI — DevDay 2026 Announcements and Developer Resources. Official announcement summary for DevDay 2026.
- OpenAI API Changelog — September 29, 2026. Technical release information for GPT-6.1 Sol, Ultrafast and Agents API computer use.
- OpenAI — Introducing the Agents API. Technical overview of hosted agent execution and the Codex harness.
- OpenAI — Sign in with ChatGPT. Identity, permission and account-sharing documentation.
- OpenAI — ZDR with Private Safety Processing. Technical documentation for privacy-preserving automated safety processing.
- OpenAI — Bedrock Managed Agents. Documentation comparing OpenAI’s Agents API with AWS-hosted managed agents.