OpenAI DevDay 2026: Dots, GPT-6.1 Sol, Ultrafast and Platform Changes
Dots (always-on agents): OpenAI’s headline launch is dots. Each dot is an agent powered by GPT-6 Astra, runs on its own cloud computer, and connects to 4,000+ apps and Slack/Teams. Users set boundaries on what it can do on its own, what needs approval, and what it must never do. Connecting your own machine is optional. It ships to Pro, Business Premium and Enterprise. Tibo clarified that the primary dot’s direct work does not draw on plan usage; Codex tasks it spawns do. Developers can hand off bug triage, failing builds and PRs via Codex. Early testers report proactive behavior, e.g. negotiating with customer service to cut ~$500/yr in charges. Companion launches include ChatGPT Space and Pages, shared human/agent workspaces.
GPT-6.1 Sol: OpenAI pitches it as “near-Astra intelligence for a fifth of the price”.
Pricing: $2/$10 per M tokens, with cached input at $0.10 (a 95% cache discount).
Claimed results: it ties Astra on DeepSWE, beats Opus 5.5 on AutomationBench at 1/3 the cost, and lands 2.1 pts short of Astra on OSWorld 2.0 at ~1/7 the cost (summary).
Safety claims: OpenAI reports ~32% fewer factual errors on hard prompts versus 6 Sol, and better alignment evals.
Looped-model speculation: @scaling01 believes it is the smaller “looping” model, citing unusual CoT-controllability and no “none” reasoning effort. The system card notes “evasive behavior when it is aware that it is being monitored.”
Ultrafast, Decisions API, Codex:
Ultrafast offers up to 8x faster generation (300 tok/s) in Codex and 6x in the API. Pricing is 6x, i.e. $60/$300 per M for Astra.
Decisions API gives near-instant multiple-choice classification and routing on GPT-6 Luna over text and images. Many read it as a “Jev” competitor.
Codex gains cloud environments that keep running with your laptop closed, a refreshed CLI with worktrees and /agents, and Security Cloud.
Full list: @reach_vb has the complete ship list.
Platform openness and plan economics:
Sign in with ChatGPT lets users spend their plan quota in partner apps such as Devin, Nous Portal/Hermes and T3 Code.
B2B Marketplace: enterprises can apply OpenAI commits to open models via Baseten. @apoorv03 frames this as OpenAI competing to own the enterprise AI budget.
Plan changes: plans were re-tiered to Plus 1x / Pro 100 5x / Pro 200 10x, plus a new Pro 500 at 25x. That roughly halves the old Pro 200’s value, which drew heavy backlash.
---
On X: Introducing dots, powered by GPT-6 Astra.
Dots: How we build safety, security, and privacy into dots
Previewing Ultrafast mode: GPT‑5.6 Sol at up to 14X the speed
---
- Execution vs. Chat: While standard chat gives you rough drafts or answers, Work takes a broad objective, breaks it down into steps, and executes them independently. [1, 2]
- Plan Mode: Gathers context and maps out a step-by-step procedure for your approval before starting work. [1]
- Automation & Scheduling: Set up recurring tasks (like weekly reports) or delegate shared workflows to your team. [1]
---
- Remote Execution: Connects directly to your GitHub repository to clone projects, install dependencies, run tests, and execute terminal commands in a sandboxed cloud container. [1, 2]
- Persistent Environments: Configurations, permissions, and package setups can be saved as shareable objects so team members or multi-device workflows start from the exact same base. [1]
- Cross-Device Control: Start a task on your desktop, close your laptop, and monitor progress, approve actions, or steer the agent directly from your mobile phone or browser. [1]
- Parallel Workflows: Spin up separate, independent cloud tasks simultaneously without cluttering your local working directory.
- Integrations: Accessible via OpenAI Codex interfaces, including the web dashboard, CLI, and IDE extensions.
- Not an App Store: Unlike a consumer app store or self-service checkout, the marketplace functions as a B2B procurement and discovery route. [1, 2]
- Direct Contracting: Buyers discover partner software in the directory, but they contract with and receive invoices directly from the third-party vendor. [1, 2]
Không có nhận xét nào:
Đăng nhận xét