Sam Altman came to Sun Valley with a simple message: the spotlight at the annual Allen & Co. Sun Valley Conference is no longer on AI promise. It is on AI costs. Speaking with CNBC, the OpenAI CEO said this was the first year AI spend has become a “very big topic” at the conference, with executives asking how to reduce spend or increase value.

And when the world’s biggest tech operators start drilling into unit economics, that is a signal boards and finance teams are already ahead of the curve. Altman said OpenAI is seeing people start to care about efficiency and getting a good return on their AI investments, and he framed it as a major consideration for the lab when developing its latest models, the GPT-5.6 family. For decision-makers, the takeaway is straightforward: AI budgets are turning into AI business cases, and model efficiency is becoming one of the levers that can make or break those cases.

The new GPT-5.6 models were released on Thursday, and Altman said they include three named variants: the flagship Sol model, the balanced everyday Terra model, and the cost-efficient Luna model. The key performance-and-cost pitch comes from GPT-5.6 Sol. Altman said GPT-5.6 Sol is “54% more token efficient on agentic coding tasks,” and OpenAI developed it with cost and speed in mind. He did not specify what the 54% figure was compared to, but he was clear about the direction: enterprises are asking for better value per token, particularly for agentic coding work where token usage can snowball quickly.

It is worth pausing on why this matters now. Altman said “every enterprise” is thinking about “spend and the value they're getting in exchange for AI.” That kind of language usually shows up when pilots turn into production, and when CFOs stop asking “can it do the job?” and start asking “can we afford to run it all day?” In practice, this is where the industry has been heading for a while, but Sun Valley is where the pressure becomes visible across the top of the ecosystem.

The attendee list this year shows why the cost conversation is accelerating. The Sun Valley conference kicked off on Tuesday and brought together major tech leaders including Netflix co-CEO Ted Sarandos, Uber CEO Dara Khosrowshahi, Apple CEO Tim Cook, and Amazon CEO Andy Jassy. When CEOs like that converge, it is rarely just networking. It is where priorities get socialized into strategy, and this year Altman’s point was that AI spend is front and center. If those companies are discussing AI efficiency at this level, the implication is that vendors and internal teams will be held to tighter performance-per-dollar expectations.

Altman’s CNBC remarks land in a broader context where enterprises are actively experimenting with tactics to stretch budgets. Business Insider notes that Altman’s comments come as enterprises consider how to achieve better returns on AI spend, and it points to a few public examples from other leaders. Coinbase CEO Brian Armstrong said in June that he was experimenting with cheaper Chinese models as defaults for his engineers and routing their prompts to the most appropriate models to avoid unnecessarily burning AI tokens. Meanwhile, Vercel CEO Guillermo Rauch told TechCrunch in an interview released earlier this week that companies need to start partnering with different AI labs for different parts of their AI stack, using models across OpenAI, Anthropic, Gemini, and Chinese players to get the best value.

Put together, these moves describe a market reality: the “single best model” approach is becoming less attractive than an orchestration mindset. Companies want the flexibility to pick or route models based on cost, latency, and task type, especially for workflows like agentic coding where inefficiency can turn into runaway compute costs. Even if executives are still excited by model capability, the conversation is shifting toward how to reliably translate capability into unit economics.

Regulators and compliance teams also sit in the same downstream reality, even if they are not explicitly named in this reporting. When AI systems become more widely deployed, finance pressure tends to increase the demand for predictable behavior, measurable outcomes, and clear governance over what models do. While this article does not describe new regulatory actions, the cost focus it captures is likely to influence procurement standards and risk appetite, because when spend is on trial, enterprises want to minimize surprises, variability, and waste.

For peers, the strategic stakes are simple and not subtle. If you are running AI at scale, the board will eventually ask whether you are buying results or buying tokens. Altman’s message from Sun Valley is that OpenAI is leaning into efficiency as a competitive requirement, not just a technical upgrade. With GPT-5.6 Sol claiming “54% more token efficient” agentic coding performance, and with OpenAI packaging cost-focused options across Sol, Terra, and Luna, the direction of travel is set: AI will be judged less by demos and more by what it costs to operate in the real world.