OpenAI DevDay 2026: Agents API and Developer Updates
OpenAI DevDay 2026 took place in San Francisco on September 29, 2026. The most important developer shift is broader than one model launch: OpenAI is bringing agent execution, tools, memory, and multi-agent coordination into managed APIs.
This guide covers the developer updates that are currently documented: the Agents API public beta, GPT-6.1 Sol, Ultrafast, and the limited-preview Decisions API, then maps each one to a practical starting point. Publication status comes from the official OpenAI DevDay page, the official OpenAI Developer Community recap, and the linked developer documentation.
What did OpenAI announce at DevDay 2026?
| Update | Current status | Why developers should care |
|---|---|---|
| Agents API | Public beta | Hosted agent execution, tools, memory, and multi-agent workflows |
| GPT-6.1 Sol | Available in the API docs | A lower-cost route for complex coding, computer use, and professional work |
| Ultrafast | GPT-6 Astra Responses API tier | Lower output-token wait time at a higher processing price |
| Decisions API | Limited preview | Classification, routing, and fixed-action decisions through a dedicated API |
| Computer use | Agents API capability | Authorized UI work in an OpenAI-hosted browser |
These labels describe different release stages. Public beta, limited preview, model availability, and access for a particular account must be checked separately in the docs, console, and request response.
1. Agents API: let OpenAI host the agent harness
Agents API entered public beta in September 2026. It gives developers a managed execution environment where agents can run code, work with files, save artifacts, and use tools across long-running tasks.
The official announcement highlights:
- a hosted sandbox for code, files, and artifacts;
- built-in tools and MCP connections;
- memory and long-running context management;
- multi-agent support that delegates independent work to parallel subagents;
- computer use in an OpenAI-hosted browser with application-controlled approvals.
The architectural change is that teams can separate model calls, tool loops, execution environments, and subagent coordination without building the entire agent harness themselves. During the beta, record tool calls, approval decisions, retries, and produced artifacts; a final chat message alone is not proof that the task completed.
2. GPT-6.1 Sol: a cost and capability tradeoff for complex work
OpenAI positions GPT-6.1 Sol as a near-Astra model for complex coding, computer use, and professional work at a lower price. The official model page lists standard text pricing of $2 / 1M input tokens, $0.10 / 1M cached input tokens, and $10 / 1M output tokens. Check the live model page for current prices and processing tiers.
Sol supports a 1.05M context window, up to 128K output tokens, and low, medium, high, xhigh, and max reasoning effort. OpenAI recommends the Responses API for tool calling; Chat Completions supports requests without tools.
Before switching, compare Astra and Sol on the same workload:
- Does a code change pass the existing tests?
- Can computer use complete the same UI steps reliably?
- How many input/output tokens, tool calls, and retries are used?
- Is any quality difference worth the lower model price?
3. Ultrafast: a faster Responses API processing tier
DevDay 2026 also introduced Ultrafast. The OpenAI Developer Changelog records Ultrafast for GPT-6 Astra in the Responses API, selected with service_tier: "ultrafast" to reduce the wait between generated output tokens. The official changelog also makes clear that regional and processing conditions apply; it is not a universal speed guarantee.
For latency-sensitive production traffic, record the model, service tier, input/output tokens, time to first token, completion time, and bill. Otherwise it is difficult to tell whether a faster result came from Ultrafast, a shorter prompt, or lower platform load.
4. Decisions API: move fixed decisions out of a general agent
The Decisions API is currently a limited preview. The official DevDay recap describes it as a specialized interface for classifying inputs, routing requests, or choosing an action from predefined answers.
Good early candidates include:
- routing support tickets into a fixed set of queues;
- deciding whether a request needs human review;
- selecting the next step from a controlled action set;
- classifying an input before handing it to a workflow.
Because it is still a preview, do not make it an unqualified production dependency. Check eligibility, response shape, version compatibility, and a safe fallback path first.
Should a new project use Responses API or Chat Completions?
Start with Responses API and Agents API documentation when the project needs OpenAI built-in tools, agent state, computer use, or multi-turn tool calls. OpenAI positions Responses API as the newer agent primitive, while Chat Completions remains useful for simple requests without tools.
| Scenario | Suggested starting point |
|---|---|
| Single-turn generation or simple classification | Chat Completions or Responses API |
| Web, file, or computer-use tools | Responses API |
| Long-running files, code, and artifacts | Agents API public beta |
| Parallel decomposition of complex work | Agents API + multi-agent |
| Fixed classes or fixed actions | Decisions API preview |
| Lower-cost complex work | GPT-6.1 Sol |
Working SDK code is only the first check. Validate tool calls, structured outputs, streaming events, approvals, and retries separately. API compatibility does not imply identical support across every model and tool.
OpenAI DevDay 2026 FAQ
When was OpenAI DevDay 2026?
The main event took place on September 29, 2026 in San Francisco. OpenAI also announced DevDay Exchanges in several cities. Check the official event page for the latest locations and dates.
Is the Agents API generally available?
No. OpenAI's announcement labels it public beta. Beta access does not mean the interface is frozen, so production systems should keep version, tool, and data migration plans.
How much cheaper is GPT-6.1 Sol than GPT-6 Astra?
The current model pages list Sol at $2 input and $10 output per million tokens, while Astra lists $10 input and $50 output. Actual task cost also depends on caching, context length, tools, and processing tier.
Will Ultrafast make every request faster?
No. It is a processing tier for eligible Responses API requests. Test it with the same model, task, and region before making a latency claim.
Can I use the Decisions API today?
The official DevDay recap currently labels it limited preview. Access depends on account eligibility and preview availability, so do not describe it as a general-availability API.
Can this site call the new OpenAI DevDay APIs directly?
Third-party gateways can expose different models, protocols, and tools from the official OpenAI platform. Check the live model pricing and API docs, then run a small request to confirm the actual model ID, endpoint, and response shape.
Sources
- Official OpenAI DevDay 2026 event page
- OpenAI DevDay 2026 announcement
- Official OpenAI Developer Community recap
- Agents API announcement
- GPT-6.1 Sol model page
- OpenAI API Changelog
- OpenAI model selection guide
Sources checked on October 1, 2026. This is a developer update and integration guide, not an independent benchmark.