OpenAI’s DevDay 2026 (Fort Mason, Sep 29, keynote 10am PT with Sam Altman, livestreamed) was not a quiet update day. OpenAI says it made “more than 20 major announcements.” The two that matter most for operators landed in the same breath: always-on agents with approval controls, and a cheaper Sol-class model aimed at agentic coding and computer use.
Dots are OpenAI’s always-on agents, powered by GPT-6 Astra, with a cloud computer, plugins, and surfaces in ChatGPT, Slack, and Teams. Custom Rules let you allow, require approval, or block actions. There is an action review path. Proactive research on connected apps is described as read-only. Rollout, per OpenAI, starts with Pro and Business Premium in eligible markets, with the first Dot at no extra cost on those plans. Enterprise, Edu, and Healthcare get a beta that admins turn on. OpenAI’s own caution is the line operators should pin to the wall: “Dots can still make mistakes, so always review consequential work.”
GPT-6.1 Sol is the other headline. OpenAI positions it as an upgrade to Sol with near-Astra agentic coding, computer use, and professional work at roughly one-fifth Astra standard token prices (OpenAI’s claim, not a Viking Labs eval). Availability today, per OpenAI: Plus, Pro, Business, Enterprise, and Edu in ChatGPT Work and Codex. Not yet in Chat. API id gpt-6.1-sol, priced at $2 / $0.10 cached / $10 per 1M input / cached / output tokens.
That pairing is the story. Always-on agents with rules. A cheaper model for the work those agents chew through. A Decisions API for finite answers when you need a router, not a novel.
Viking Labs’ take stays the same: route by job, keep humans on judgment, and only automate inside bounded workflows.
What else OpenAI shipped (attribute to OpenAI)
Beyond Dots and Sol, the DevDay 2026 recap and companion posts cover a long tray of ships. Operators should skim these with a job map, not a FOMO list:
- Ultrafast. OpenAI claims up to 8× on Codex and 6× on the API (about 300 tokens/s). GPT-6 Astra Ultrafast is available today on the API and on Work/Codex for Pro 500 and Enterprise. Sol Ultrafast is listed as coming soon. Among Pro plans, Ultrafast is only on Pro 500.
- Pro 500. A new Pro tier with 25× Plus usage plus Ultrafast, available now. OpenAI Help Center lists Pro 500 at $500/month (About ChatGPT Pro tiers). Same Help Center table: Pro 100 is $100/month; Pro 200 is $200/month and is available again for new subscribers, with a lower included usage allowance for new non-grandfathered plans. Eligible existing Pro 200 subscribers keep their prior allowance through Oct 29, 2026, then move to the lower allowance. Keeping the grandfathered allowance does not add Ultrafast or Pro 500 features.
- Decisions API (Luna). Finite answers for routing-shaped work, limited preview. Useful when the job is “pick one of N,” not “write me an essay.”
- Codex stack. Codex in the cloud (run on computer, phone, or cloud; reusable envs), a refreshed Codex CLI (voice,
/agents, worktrees), Code Review on ChatGPT desktop, and Codex Security Cloud (scan GitHub repos on demand or schedule; investigate, dedupe, prepare fixes). - Agents API and computer use. Continues the agent runtime surface OpenAI introduced earlier in September (Agents API post). Bedrock Managed Agents with Amazon appear in the same DevDay wave.
- ChatGPT Space, Pages, and collaboration. Space and Pages land now. Collaborative slides are listed as coming in weeks. Teams get shared and recurring tasks. @ChatGPT lands in Slack and Teams. Meetings plugin beta (macOS).
- Identity and ecosystem. Shareable profiles, Sign in with ChatGPT (16 partners named by OpenAI), Marketplace (32 partners), plus plugin extensions, creator tools, Sites, and MCP events.
- Private Intelligence / Private Inference. Preview expected in the fall, aimed at private inference paths.
Full primary writeups: Dots, GPT-6.1 Sol, DevDay recap, Pro tiers Help Center, event hub.
The operator filter (not the keynote filter)
Vendor days reward volume. Ops rewards clarity. Ask three questions before you turn any of this on:
- What is the job? Research draft, ticket triage, dashboard refresh, code review note, meeting follow-up. Name one job. Not twelve.
- Where does a human still own the gate? Sends, deletes, money moves, customer-facing claims, anything irreversible. Custom Rules and action review only help if you write the rule before the agent gets keys.
- Is the workflow bounded? Inputs, tools, success metric, and three exits: skip, ask a human, refuse. If you cannot draw that box, you do not have automation. You have a demo with privileges.
Route by job. A Decisions API-shaped path for finite routing. A Sol-class model for agentic coding and computer use when price and speed matter. Astra-class power when the job earns it. Dots for always-on work that still stops at your approval line.
Do not confuse “always-on” with “always-trusted.” OpenAI said Dots can still make mistakes. Believe them.
What Viking Labs builds
We are a software development company. We build SaaS (including financial data dashboards), productized automations, and done-for-you agent builds with hard scopes and approval gates. Custom software when automation is not enough. ResearcherFlow is one production proof we run ourselves: live permissions, billing, and human gates on writes. Evidence, not the whole catalog.
If you want the short tour: Services, Products, Work with us, About, Newsletter.
Start with one bounded workflow
DevDay 2026 gave operators more always-on surface area and a cheaper Sol path. The judgment gate did not move. It still sits with the human who owns the write.
Bring one job. Clear tools. A human gate. A way to measure hours returned. We will tell you whether that job wants a productized automation, a custom SaaS surface, or a done-for-you agent build. Sometimes the honest answer is “not yet.” That is still useful.
Jon Marrs · About Viking Labs
Leave a Reply