OpenAI’s DevDay 2026 landed several developer-facing changes at once: the Agents API hit public beta with computer use support on September 10, GPT-6.1 Sol arrived at one-fifth of Astra’s price on September 29, and the Decisions API entered limited preview for faster classification and routing. Not everything is ready to drop into your stack today — here is what is actionable, what needs a waitlist, and what is just product theater.
Agents API: Computer Use Is in Public Beta #
The Agents API now lets you build agents that operate software through graphical interfaces. You declare tools=[{"type": "computer"}] in the Responses API, and the model returns a computer_call with a batched actions[] array: click, double_click, drag, move, scroll, keypress, type, wait, and screenshot. OpenAI runs the underlying VM — you do not manage execution infrastructure yourself.
Access is available via the API directly, through Codex, and inside ChatGPT Work on the Pro $500 and Enterprise plans. Full documentation is at the OpenAI computer use guide.
To be direct: this is not a capability lead. Anthropic shipped computer use in late 2024. Google Gemini followed in September 2026. OpenAI is third to market here. The real differentiator is the hosted infrastructure model — OpenAI manages the execution environment — and the AWS integration described below.
For teams on AWS, Amazon opened a public preview of Bedrock Managed Agents powered by OpenAI on the same day. Agents sign in with AWS IAM identities instead of an OpenAI API key, which means your existing permission models, audit logs, and compliance tooling all apply. Available now in US East (N. Virginia), US West (Oregon), and US East (Ohio) with no service charge during preview.
GPT-6.1 Sol: The Pricing Is the Story #
GPT-6.1 Sol is priced at $2 per million input tokens and $10 per million output tokens — exactly one-fifth of GPT-6 Astra ($10/$50). The cached input price is $0.10 per million, which is 95% below standard input and the sleeper number in the entire announcement.
| Model | Input $/M | Output $/M | Cached $/M |
|---|---|---|---|
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| GPT-6 Astra | $10 | $50 | — |
Performance holds up. On DeepSWE v1.1, Sol matches Astra. On OSWorld 2.0, it comes within 2.1 points of Astra at roughly one-seventh the cost per task. On Terminal-Bench Science, Sol averages $5.47 per task versus $23.80 for Astra. OpenAI recommends Astra for the most difficult research tasks; everything else is fair game for Sol.
The action item is straightforward: audit your current model usage. If you are running agentic coding workflows on Astra, moving to Sol is a legitimate 5x cost reduction with minimal quality trade-off on most workloads. The $0.10 cached input price makes prompt caching nearly free — worth implementing if you have not already. The Next Web has a detailed pricing breakdown if you want to run the numbers for your specific usage.
Decisions API: Compelling, But Not Available Yet #
The Decisions API is in limited preview and introduces something genuinely new: a constrained-output endpoint powered by GPT-6 Luna that returns only one of your predefined answers, not a generated paragraph. You supply context (text or images) and a finite answer space; the API returns a decision.
OpenAI claims 150ms latency versus 1.6 seconds for a standard Luna call — a 10x speedup that matters for real-time routing in agent loops. Use cases include content classification, request routing, and choosing the next step in a multi-agent workflow.
The honest caveat: pricing, rate limits, regional availability, and maximum choices per request are all unpublished. You can request access, but you cannot model production costs yet. Keep an eye on the official DevDay recap for updates — OpenAI said broad release was coming in the days following the event.
Dots and ChatGPT Space: Watch, Do Not Build On Yet #
Dots are persistent agents built into ChatGPT, bundled with the $100/month Pro plan. ChatGPT Space is the new collaborative workspace for Pro, Business, and Enterprise users. Both are product-layer features — neither exposes a third-party API today. They signal where OpenAI is taking the product, but they are not yet actionable for external developers. The InfoQ DevDay breakdown covers these features in more depth if you want the product context.
Three Things to Do This Week #
- Enable computer use in the Agents API — public beta is live; add
tools=[{"type": "computer"}]and test your use case against the docs - Audit model costs against Sol pricing — most agentic coding workloads can shift from Astra to Sol with a straightforward model swap and a 5x cost reduction
- Apply for Decisions API preview — if you build routing or classification layers, get on the list before pricing details land