Building Abundant Intelligence

OpenAI presented an 'abundant intelligence' strategy to make advanced AI cheaper and more widely usable. The idea is that as the cost of useful intelligence falls, more tasks become economically viable, and expanded usage in turn feeds back into research, infrastructure investment, and model improvements.

With a recent price adjustment, GPT-5.6 Luna's price dropped 80% to $0.20 per 1 million input tokens and $1.20 per 1 million output tokens. GPT-5.6 Terra was cut 20% to $2 for input and $12 for output, while Fast mode for GPT-5.6 Sol runs up to 2.5x faster than standard processing but costs twice as much.

OpenAI explained that what matters is not simple token pricing but the total cost of achieving a successful outcome. When accounting for retries, human oversight, and error correction, a more powerful model can be more economical than a cheaper one, while conversely, lower-cost models can maintain the same quality while broadening AI accessibility.

To improve compute efficiency, OpenAI is improving not only models but the overall system. GPT-5.6 Sol optimized the operational software used to serve models, cutting overall serving costs by 20%, and improvements to speculative decoding raised token generation efficiency by more than 15%.

Routing, context management, tools, and product design were also cited as key factors. In particular, improvements to retained reasoning and context management raised the score on the public ARC-AGI-3 benchmark from 13.3% to 38.3% without any model change, while reducing output token usage to one-sixth.

OpenAI emphasized that its full-stack approach—building infrastructure, models, platforms, and products all together—lets each layer improve the others. Real usage data and demand from ChatGPT, ChatGPT Work, Codex, and the API inform research direction and capital investment decisions, and the resulting research improvements in turn boost product performance and lower service costs.