Build Hour: Valuemaxxing with GPT-5.6

### TL;DR OpenAI's Build Hour introduces "valuemaxxing" with GPT-5.6: shift from measuring AI consumption (token count) to measuring outcomes, and use models like Soul, Terra, Luna with smarter reasoning to get more done with fewer (or more) tokens when it truly adds value. ## From Token Maximization to Value Maximization Christine (startup marketing) and Charlie (developer experience) kick off this Build Hour with a *The Office* meme: "Getting more value out of fewer tokens." After launching GPT-5.6, they asked X (formerly Twitter) why users loved it. Common theme: remarkable token efficiency — "doing more with less." Two days before the event, they hit a new milestone on X, thrilled that the model resonates with people who want to maximize output per token. Token maximization (earlier this year) meant measuring progress by AI usage: how many tokens consumed, how many prompts sent, how many agents managed. Some companies built internal leaderboards tracking token usage per employee per week. But sentiment shifted: some companies burned their full-year AI budget in months; others pulled back, saying "maybe we should cap AI spending"; even some AI CEOs noted that companies might be too obsessed with token leaderboards. The important thing became: measure by output and value created. This leads to **valuemaxxing**. Instead of how much AI you used, it’s what AI helped you *accomplish* — work completed, time saved, quality of code or deliverables. Charlie’s framework: ask yourself, "If your token spend doubled tomorrow, how would you know if it was worth it?" Not naively assuming more tokens = better. Questions to ask: - What outcomes do you want to improve? - What workflows produce high-quality results for those outcomes? - Where can you inject intelligence, AI, tokens to accelerate or improve quality? - What evidence shows you're improving quality or building an effective system? In AI collaboration, you’ve heard "evals" — how to define "good." As we move from simple chat tasks to AI agents owning whole outcomes, defining and tracking "good" for your company or role remains essential. Sometimes the answer is "spend fewer tokens" (cut costs). Other times, spend *more* tokens to drive more value: - Ensure high quality → increase model reasoning or switch to a larger model. - Save real time by spending more tokens to go faster. - Accelerate a large migration project: spend more tokens to finish in weeks instead of months. - Scale a working workflow: spend more to expand output. - Manage risk with a high-quality, bulletproof system using LLM-as-judge with more or diverse models to catch edge cases. ## Valuemaxxing with Codex Two lenses: individual developer using **Codex**, and organization building on OpenAI API. Within Codex, the story starts with the **GPT-5.6** series (three new models): - **Soul** – flagship, most capable, for complex coding and specialized tasks. - **Terra** – everyday workhorse, balancing intelligence with cost/latency. - **Luna** – high throughput, low intelligence requirement, focused on cost and latency. OpenAI invested heavily in token efficiency for these models. On a popular coding benchmark (Deep Suey), the 5.6 series shows best price/performance (value per token). But importantly: a model that isn't smart enough in one shot can reach similar performance if you are willing to spend *more tokens*. So evaluate total cost to complete a task, not just cost per token. ### Interactive Demo: Valuemaxxing Workbench Charlie built a demo with GPT-5.6 and Codex (code to be open‑sourced after the Build Hour). First page: a 3D chart comparing leading models on intelligence, cost, and speed. More interesting: a visualization matrix where, with API keys, you can render SVG outputs for different models and reasoning levels. Example prompt: "A panda riding a jet ski" instead of "A pelican riding a bicycle." Running the matrix shows how quality varies. With Luna, first result rough, but with more reasoning, it becomes decent. With Soul, base fidelity high, gets even stronger with added reasoning. ### Practical Tips for Using Codex 1. **Start with Soul “medium” reasoning.** Many developers default every completion to “extreme high.” For complex, deep projects, yes; but many daily tasks work well with Soul medium. If it’s not smart enough, crank up. If it’s not coding‑heavy, drop to Terra. 2. **Consider when to trade tokens for time.** Sometimes you want faster output and are willing to spend more tokens. Example: Turbo mode (1.5x speed) but uses more quota. **Auto‑approve** is highly recommended as default (not full access – that’s too permissive, unless you have a separate model reviewing output). Auto‑approve is a good compromise. **Chronicle** (new in Codex): records your screen and builds memory for tasks. It costs more tokens but gradually learns your daily workflows. 3. **Audit your `agent.md` and skill instructions.** Models are now smarter. Many explicit instructions written 6–12 months ago may be unnecessary. Review instruction‑following and persistence: do you really need everything spelled out? Can you compress to save input tokens? ## Optimizing Your Toolchain (API Features for Products) Moving to building products on the OpenAI API. Recently introduced features that help developers get more value: - **Programmatic Tool Calling** – a significant new feature. (Transcript cuts off here; the recording likely continued with more details.) ## Upcoming: Floy’s Migration Story Christine mentioned a segment where Floy (a promising startup) would share how they migrated their AI agent. (This part was not included in the provided transcript, so no details are available.) ## Q&A The session was intended to include a Q&A segment where the team answers questions submitted via the Q&A button. (Not transcribed.) --- **Source:** [Build Hour: Valuemaxxing with GPT-5.6 – OpenAI YouTube](https://www.youtube.com/watch?v=jyuyY86GJnA)