Cheaper Tokens Fuel Always-On Agents

Lower token prices push model usage away from one shot chat and toward always on software, which can expand OpenAI’s API revenue even as each individual token gets cheaper. Cheap models make it practical to run coding agents that iterate through files, tests, and fixes, and to keep background jobs running on schedules or long lived workflows. Those products burn far more tokens than a single prompt and reply, especially when they repeatedly call models around the clock.1 2 3 4 5 6

• Coding agents are a concrete example of token volume expansion. OpenAI launched Codex CLI in April 2025 and the Codex cloud agent in May 2025, then reported daily Codex usage grew more than 10x since early August 2025, with GPT-5-Codex serving over 40 trillion tokens in three weeks after launch.2 3 7
• Background work changes the usage pattern. OpenAI’s API supports running responses with background set to true, which fits jobs that continue after the user leaves, while prompt caching lets repeated instructions and context be reused at discounted rates. That makes recurring automations cheaper to operate and easier to leave on continuously.4 5 6
• The pricing curve is steep enough to unlock new use cases. OpenAI’s current API pricing lists GPT-6 Sol at $2 input and $10 output per 1M tokens for standard short context, versus GPT-4 launch pricing of $30 input and $60 output per 1M tokens. Sacra estimates OpenAI’s annualized revenue reached $50B by September 30, 2026, showing lower unit pricing has coincided with much larger overall demand.1 8

This points toward an API business driven less by premium pricing on isolated calls and more by high frequency machine labor embedded in products. As coding copilots, autonomous workflows, and agent platforms become routine software infrastructure, the winners should be the labs that turn lower prices into much higher task volume and habitual background consumption.2 4