A set of numbers ended the "agents are still early" line
At DevDay, OpenAI dropped a set of figures: ChatGPT passed 1.2 billion weekly active users, Codex and ChatGPT Work together passed 35 million weekly users, and 2.5 million businesses use OpenAI products. The annualized revenue run rate is approaching 70 billion dollars, up about 70 percent since the start of the third quarter.
1.2 billion weekly users means roughly one in seven humans touches ChatGPT each week. But the number to watch is 35 million for Work and Codex, because it represents not just chatting but people actually handing workflows to agents.
The price war is pulling unit cost down
At the same DevDay, OpenAI launched GPT-6.1 Sol, which it says approaches Astra's capability at one-fifth the cost. This is not an isolated move; the whole frontier is competing on "good enough and cheap enough." Anthropic's annualized revenue passed 65 billion dollars back in July, its IPO may land as early as October, and investors are talking up a 2 trillion valuation. On the voice side, ElevenLabs just completed a 300 million dollar employee tender at a 22 billion dollar valuation, and its ElevenAgents already make up 55 percent of revenue.
Put these together: model capability is approaching a ceiling, but the price of a unit of intelligence is dropping fast. The "hourly rate" of agent labor is being repriced.
But weekly users are not economic value
Dialectically, inside those 35 million weekly users, some only occasionally ask an agent to look something up, while others have delegated recurring operations entirely. Both count as weekly users, yet the economic value they create differs by orders of magnitude. Frequency, task depth, and whether you reach the "autonomous execution" step are the real dividing lines.
Another overlooked point: multi-model adoption is weakening any single vendor's lock-in. One organization may use OpenAI for general assistance, Anthropic for coding, and another provider for specialized loads. Once interfaces and integrations standardize, buyers can shift loads to the cheaper or more reliable vendor at will. So-called lock-in is not that solid.
The voice line deserves a separate look
ElevenLabs at a 22 billion valuation with ElevenAgents over half of revenue shows that "an agent that can talk and make calls" already supports a large standalone business. When voice synthesis is nearly free and cloning takes seconds, voice labor such as support, outbound calls, and companionship gets replaced by agents first, not because the models are strongest, but because that line's unit cost is collapsing fastest.
The consumer front door became an enterprise sales funnel
OpenAI's play is clear: let over a billion people get fluent in personal contexts, then when they enter a company, procurement is just a formal license for a tool they already use. This consumer-habit-to-enterprise pipeline beats any sales army on cost. For competitors and small teams, you must compete at the front-end experience, not only on API price, because once the habit forms switching cost is brutally high.
The small-business window is opening
Falling unit-intelligence prices mean small businesses can afford automation that once only large firms could sustain. But do not hand core flows to agents on day one. Pick a clearly bounded, reversible task first, such as feedback classification, weekly reports, or competitor monitoring, and expand after it works. What separates winners is which slice of work you dare to run automatically, not how many trends you chased.
What this means for you
For a small business like yours, the conclusion is concrete: the unit cost of agent labor is dropping fast, and seriously considering adoption is no longer "following a trend" but an accounting question. Start with a clearly bounded, reversible task, such as classifying customer feedback or auto-generating weekly reports, rather than waiting for a perfect plan. But remember, most of those 35 million are shallow users; what pays is which slice of work you dare to let it run automatically.