What OpenAI DevDay 2026 Means if You Actually Ship Software
Always-on agents got the keynote. For builders, the important announcements were a $2 model, a managed agents API and a release OpenAI pulled.
OpenAI held DevDay 2026 on Tuesday 29 September. The keynote centred on dots, always-on agents inside ChatGPT that run on GPT-6 Astra, connect to more than 4,000 apps and can be reached through Slack and Microsoft Teams. They come with read-only proactive research and custom approval rules, and are limited to the top paid tiers and admin-enabled business workspaces for now.
Dots will get the headlines. As someone who builds products for clients in Malaysia, I care more about what changes in the stack I pay for. Here is my read.
GPT-6.1 Sol is the default to test
A week after GPT-6 Sol, OpenAI shipped GPT-6.1 Sol. It costs $2 per million input tokens and $10 per million output tokens, against $10 and $50 for Astra, and cached input drops to $0.10 per million, a 95 per cent discount on its own input price, iThinkDiff reported. OpenAI claims it matches Astra on DeepSWE v1.1 and trails it by about two points on OSWorld 2.0 for computer use. It is in ChatGPT Work and Codex now, with the API name gpt-6.1-sol.
Two things stand out for anyone running agents. First, the cache price. Agents resend the same system prompt, tool schemas and history on every step, and that repeated context is usually most of the bill. Second, the price point. Anthropic's Claude Sonnet 5.5, released the day before DevDay, lists at the same $2 and $10 per million tokens with a one-million-token context. Two serious mid-tier models at an identical price means you can choose on quality for your task, not on cost, and switch if one regresses.
A managed agents API, and a narrower one
OpenAI also launched an Agents API with computer use, multi-agent support, tool search and context compaction, running on infrastructure OpenAI manages, according to Reworked's rundown. That removes a lot of plumbing that small teams currently build themselves. It also deepens lock-in: if your agent loop, sandbox and memory live in OpenAI's cloud, moving to another provider later becomes a rewrite.
The quieter launch is the one I expect to use: a Decisions API, in limited preview, that uses GPT-6 Luna to answer predefined questions with a fixed set of answers. Classification, routing and choosing a next step are most of what production AI features actually do. A purpose-built endpoint for constrained decisions should be cheaper and easier to test than a general chat model asked to reply in JSON.
The pricing shuffle
ChatGPT gets a $500-a-month Pro tier with 25 times the Plus allowance, and an Ultrafast option that OpenAI says generates up to eight times faster in Codex and six times faster in the API, at six times standard API pricing. Meanwhile the $200 Pro tier drops from 20 to 10 times the Plus allowance for new subscribers, BGR noted. For a team in Kuala Lumpur paying in ringgit, that is a real cost increase dressed up as a new option. Check your seat plans before renewal.
The model that did not ship
The most telling news came just before the keynote. OpenAI cancelled the planned release of GPT-6.1 Astra after internal testing found it was less honest about which actions it had taken and would carry on with tasks, and reach for outside tools, without the user's permission, Android Authority reported, citing The Wall Street Journal's interview with Saachi Jain, OpenAI's head of safety systems. OpenAI says it will focus on making future models safer, and the model may still ship later.
Credit to OpenAI for pulling it. But note what the failure was: acting beyond its authorisation. That is the exact behaviour that makes always-on agents risky, and it showed up in the model one step up from the one powering dots. If you are building agent features, scope permissions tightly, require confirmation for anything irreversible, and log actions rather than trusting the model's account of them.
My short version: test GPT-6.1 Sol and Sonnet 5.5 side by side on your own tasks this week, try the Decisions API if you get access, and be cautious about putting your whole agent loop in anyone's managed cloud.
Sources
- Everything OpenAI Announced At DevDay 2026 — BGR
- The biggest announcements from OpenAI DevDay — Reworked
- OpenAI Launches GPT-6.1 Sol with Near-Astra Performance at One-Fifth the Price — iThinkDiff
- OpenAI cancels GPT-6.1 Astra over safety concerns — Android Authority
- Models overview (Claude Sonnet 5.5 pricing) — Claude Platform docs