OpenAI CEO Sam Altman states that AI work assistants may only need one more model iteration to become very useful
OpenAI CEO Sam Altman stated that AI work assistants may only need one more model iteration to become very useful.
He emphasized the distinction between an "assistant" and a chatbot: existing systems can write drafts, check emails, and connect with Slack and calendars, but they still require repeated prompts and revert to tool status when context is lost. The next generation is described as capable of integrating emails, texts, documents, and meeting audio, continuously observing screens and workflows within user authorization, directly producing sales emails, strategy documents, and scheduling conflicts without asking, "What do you want me to do?".
The product side corresponds to the launched ChatGPT Work: combining ChatGPT with Codex capabilities at the same workstation, reading user-selected applications, documents, and processes, outputting documents, spreadsheets, presentations, and small web applications, and integrating with Gmail, Drive, calendars, and CRM. Internal demonstrations included generating charts from Slack engineering discussions and transferring a disorganized academic calendar from email to calendar. Altman himself has used an agent to handle a large volume of morning messages and described the experience of such systems executing tasks across applications as "magical".
He previously used a more stringent analogy for capability levels: today's agents are like interns who can work for a few hours, while the next stage should resemble skilled engineers who can work continuously for several days; some programmers have reported efficiency improvements of two to three times, and he anticipates a scenario where "one person plus a model can replace an entire software team." GPT-6 Astra is referred to as the strongest model for computer usage, professional work, science, programming, and cybersecurity, aiming to enable agents to operate autonomously on computers for extended periods.
Competition is simultaneously pressing within the same generational gap. Google is testing all-day personal agents within the Gemini system, covering emails, documents, and calendars; Anthropic is pursuing Claude Cowork, and Perplexity is exploring browsing agents. OpenAI has also recruited Peter Steinberger, founder of OpenClaw, betting that the next generation of personal agents will quickly become core products. The bottleneck is no longer just benchmark scores, but rather the scope of authorization, continuous memory, and risks of misuse: once an assistant is allowed to view the entire screen and inbox, its utility and potential for data leakage both increase.
In market mechanisms, this is a seat competition driven by model iteration, not a one-time feature release. Buyers are enterprises subscribing to compress white-collar processes and high-value individual seats, while sellers are laboratories that can securely integrate the next generation of models into existing SaaS permission systems. Funding flows towards computation, security audits, and plugin directories; beneficiaries are platforms that first present "next generation" as a deliverable timeline, while pressured parties are tool vendors still relying on per-prompt licensing sales and unable to enter continuous workflows, as well as enterprise data rooms with overly broad permissions.
Source: Public Information
ABAB AI Insight
Altman is not defining a new category for work assistants, but rather pushing forward the "super capable colleague" he mentioned in 2024: that colleague should know all emails and conversations without making people feel possessed. OpenAI has progressed through ChatGPT Agent, Operator, Deep Research, and now ChatGPT Work, following the same path—first enabling the model to use a computer, then allowing it to see your computer. The recruitment of Peter Steinberger and OpenClaw indicates the laboratory's judgment: the breakthrough point for personal agents lies in execution, not in another dialogue box.
The capital path is subscription seats multiplied by permission depth. The $20 tier has already been integrated into Work, and enterprise tiers will further sell connections to Slack, email, and CRM. The motivation is to shift token consumption from Q&A to all-day operation: the more hours the assistant spends observing the screen, the higher the reasoning costs and data retention, also locking in user migration costs within its own plugin directory. Microsoft Copilot has already occupied the office suite, Google uses Gmail and Docs as default entry points, and OpenAI must leverage the next generation of models to make "completing full tasks" a reason for platform switching.
The analogy is the leap of smartphone assistants from voice commands to system-level permissions, similar to Salesforce transforming CRM from a recording system to an action system. The industry is transitioning from tool assistance to controlling workflows: whoever first enables the assistant to complete tasks without prompts will shift from being a model supplier to an operational system for workstations. Expansion continues, with control shifting from parameter quantity to permission mapping.
Structural judgment belongs to the pricing power shift driven by technological substitution. The mechanism is that white-collar time is divided into automatable hour packages: drafts, retrievals, scheduling, and cross-application transfers are first consumed, while judgment and accountability remain with individuals. Once the next generation of models surpasses "very useful," pricing power will shift from per-word charges to charges based on agent working hours and permission packages, with entry-level positions on the labor cost sheet being rewritten first.
ABAB News · Cognitive Laws
Tools require people to ask, assistants need to observe themselves
The value of the next generation model lies in the permissions it is allowed to access
Whoever first completes the entire task will redefine workstation pricing