OpenAI has released GPT-6 Sol and GPT-6 Luna, turning model cost into one of the central parts of its latest product push. The two models are positioned as faster and more affordable members of the GPT-6 family, building on the earlier GPT-6 Astra model while targeting workloads where customers need large volumes of inference.
The timing is important. AI companies are increasingly competing not only on what their models can accomplish, but on how much computation is required to accomplish it. A model that produces a slightly better answer but consumes dramatically more tokens can be difficult to deploy at scale. OpenAI’s latest release So puts efficiency alongside capability.
- The company says GPT-6 Sol and Luna are rolling out across ChatGPT Work and Codex and are also available through the API. OpenAI’s API documentation lists standard pricing of $2 per million input tokens and $10 per million output tokens for GPT-6 Sol, while GPT-6 Luna is listed at $0.10 for input and $0.50 for output under standard processing.
- Those prices do not tell the whole story. Enterprise AI bills depend on caching, context length, tool calls, retries and the number of steps an agent takes to finish a job. A cheaper model can still become expensive if it needs many more calls. Conversely, a stronger model can reduce total cost if it completes a task in fewer attempts.
- This is where agentic software is changing the model market. Developers increasingly route different parts of a workflow to different models. A difficult reasoning step may use a large model, while file classification, extraction or routine changeations can run on a cheaper one. Model providers So have an incentive to make the lower-cost tiers useful enough to carry a large share of production traffic.
OpenAI’s move also arrives as Anthropic cuts the operating cost of Claude Opus 5.5 and as Google continues expanding its own accelerator infrastructure. The competition is becoming a three-way race involving model quality, inference efficiency and the hardware underneath it.
For businesses, the practical question is no longer simply which model has the highest benchmark score. It is how much useful work the model produces per dollar, how reliably it follows a workflow and how much infrastructure is required around it. GPT-6 Sol and Luna show that OpenAI is treating that calculation as a core part of the product.
That makes the next few months more revealing than the launch itself: adoption, reliability and cost will determine whether the technology becomes routine infrastructure.
OpenAI is also making the model-selection problem more complicated by provideing multiple levels of capability. That can be useful for developers building systems that route work according to difficulty. A production application can reserve the expensive model for the steps that need it and use the cheaper model for routine work, potentially changing the economics of an entire workflow rather than one API request.
For developers, predictable cost may ultimately matter as much as another benchmark point.
OpenAI’s pricing move also changes the competitive pressure on smaller model providers. If a large lab can provide a relatively inexpensive model with strong reasoning, an open-weight or specialist model has to compete on something other than price alone. That could push the market toward specialization, local deployment and models improved for specific industries rather than a single general-purpose winner.
That makes the next phase of the model race a question of economics as much as intelligence.
For OpenAI, the lower-cost models also create room for applications that would have been too expensive at earlier prices. Coding agents, document processing and background automation can make many small model calls rather than one large request. If those calls become cheap enough, developers can build systems that continuously inspect, classify and update information instead of waiting for a person to trigger every operation. That is one reason inference pricing has become a product story in its own right.