OpenAI Explains Value Maximization with GPT-5.6

OpenAI recently hosted an "OpenAI Build Hour" session focused on the concept of "Value Maximizing with GPT-5.6." The session featured Christine from OpenAI's startup marketing team and Charlie from the developer experience team, who were joined by Bryant Chou and Lorenzo Gentile from the startup Ploy. The discussion aimed to guide developers on how to extract the most value from OpenAI's APIs and models, moving beyond simply maximizing token usage.

From Token Maximization to Value Maximization

Charlie began by explaining the shift in mindset from "token maxing," which he defined as measuring progress by the sheer volume of AI usage (tokens burned, prompts sent, agents managed), to "value maxing." This new approach emphasizes measuring progress by what AI helps accomplish, such as work completed, time saved, and quality improvements.

The full discussion can be found on OpenAI Youtube's YouTube channel.

He posed a critical question for developers to consider: "If you doubled your token spend tomorrow, how would you know if it was worth it?" This prompts a deeper analysis of desired outcomes, workflows where AI can add intelligence, and the evidence of actual quality improvements.

Leveraging GPT-5.6 Models and Codex

The session highlighted OpenAI's new GPT-5.6 family of models: Sol, Terra, and Luna. Sol is positioned as the frontier model for complex tasks, Terra as a balanced daily driver, and Luna for high-volume workloads where cost and latency are paramount. The presentation emphasized the token efficiency of the GPT-5.6 family, citing benchmarks like DeepSuite to demonstrate improved performance per token spent.

Charlie showcased a "value maxing lab" built with GPT-5.6 and Codex, which visualized model performance against cost and speed. This demo allowed for a visual comparison of different models and reasoning levels, illustrating how increased token expenditure on higher reasoning levels can yield better results.

Codex Tips and Tricks for Value Maximization

Practical advice was offered for developers using Codex:

• Start with Sol on Medium: This model is recommended as a strong starting point for many day-to-day coding tasks.
• Trade Tokens for Time: Features like "Fast Mode" can accelerate output at the cost of increased token usage. Auto-approval is recommended over full access for safety and efficiency.
• Audit AGENTS.md and Skills: With more sophisticated models, older instruction-following or persistence mechanisms might be redundant, allowing for more concise prompts and reduced token input.

Optimizing Harnesses with API Features

For organizations building on OpenAI's APIs, several key features were highlighted for value maximization:

• Programmatic Tool Calling: This feature allows models to write and execute JavaScript code in a sandbox, moving complex computations out of the model's reasoning chain and saving tokens and time.
• Prompt Caching: Developers can specify which parts of prompts to cache, saving costs on repeated inputs. Appending dynamic information (like dates) to the end of prompts is recommended to maximize cacheability.
• Persistent Reasoning: This feature saves reasoning turns across conversations, improving continuity and cache efficiency for chained assistant messages.
• Compaction: This feature helps reduce the amount of input sent by compressing long histories and tool traces, making extended workflows more efficient.

A demo was presented showcasing A/B tests on these features, demonstrating significant input cost reductions (e.g., 90% lower with prompt caching) and improved efficiency.

Customer Spotlight: Ploy's Migration to GPT-5.6

Bryant Chou and Lorenzo Gentile from Ploy shared their experience migrating their AI agent to GPT-5.6 Sol. Ploy is a marketing platform that helps businesses turn their websites into growth engines by managing SEO, AEO optimization, conversion improvements, and ad campaigns.

They highlighted that their AI agent handles tasks like de-anonymizing website visitors, qualifying leads, syncing with CRMs, and drafting emails. The migration to GPT-5.6 was smooth, with their prior optimization work, particularly around caching, carrying over effectively. They emphasized the importance of caching for cost efficiency and responsiveness, noting that proper implementation is crucial to avoid breaking the cache. They also shared strategies for loading tools on demand without breaking the cache and using breakpoints to optimize cache hit rates.

The session concluded with a Q&A, addressing topics such as the impact of agent depth on performance and instances where starting a new chat might be preferable to compaction.