Build Hour: Valuemaxxing with GPT-5.6

This OpenAI Build Hour session, led by Charlie Guo (DevEx) and featuring Bryant Chou (Founder, Ploy) and Lorenzo Gentile (AI Engineer, Ploy), focuses on “value maxing” with the new GPT-5.6 models, empowering developers to achieve more useful work from every token.

The Shift: From Token Maxing to Value Maxing

The session introduces a critical paradigm shift from “token maxing” to “value maxing”.

• Token Maxing: Initially, progress was measured by AI usage metrics like tokens burned, prompts sent, or agents managed, sometimes even with internal leaderboards. However, this led to unsustainable spend and a realization that output-based measurement was crucial.
• Value Maxing: This approach measures progress by what AI accomplishes – work done, time saved, quality improved. The core question is: “If you doubled your token spend tomorrow, how would you know if it was worth it?”
  • Self-reflection questions for companies: What outcomes need improvement? Which workflows lead to good outputs? Where can AI add intelligence? What evidence proves quality improvement?
  • Importance of Evals: Defining “what good looks like” is paramount, especially as AI agents become responsible for entire outcomes.
  • Strategic Token Spending: While some cases might warrant spending fewer tokens, there are scenarios where more tokens drive greater value:
    • Ensuring high-quality results by using larger models or higher reasoning.
    • Achieving faster completion (trading tokens for real-world time).
    • Scaling successful workflows.
    • Managing risk through diverse LLM judges.

Value Maxing as an Individual Developer (CodeX)

The GPT-5.6 family (Sol, Terra, Luna) is central to token efficiency:

• GPT-5.6 Sol: Frontier model for complex coding and professional tasks.
• GPT-5.6 Terra: Balanced daily driver for intelligence, cost, and latency.
• GPT-5.6 Luna: High-volume workloads where cost and latency are critical, and intelligence demands are lower. GPT-5.6 models offer the “most bang for your token” across various reasoning levels, emphasizing “cost per overall task to completion” rather than just “cost per token.”

CodeX Practical Tips:

1. Start with Sol on Medium: This is a strong default. Tune up if more intelligence is needed, or down to Terra/light reasoning for less intensive tasks.
2. Trade Tokens for Time:
   • Fast Mode: Runs outputs at 1.5x speed (consumes usage faster).
   • Auto-approval: Recommended for efficient workflow, balancing safety without constant permission requests.
   • Chronicle: Records screen activity, developing task memory for repeated similar tasks, which uses more tokens but improves efficiency.
3. Audit agents.md and Skills: Models are more sophisticated; much of what was explicit 6-12 months ago might now be redundant. Make instructions concise to save input tokens.

Value Maxing Your Harness (OpenAI APIs)

Several API features help developers get more value:

1. Programmatic Tool Calling:
   • Introduced with GPT-5.6, this allows the model to write code for a JavaScript sandbox.
   • The sandbox executes tool calls and computations, moving this work out of the model’s chain of thought.
   • Benefits: Saves tokens (less model reasoning), saves time (batch tool calls without round-trips).
   • Demo Result: 24% fewer input tokens, saved an entire model turn.
2. Prompt Caching (KV Caching):
   • Ability to specify which parts of prompts to cache.
   • Benefits: Saves time and tokens, especially for large, unchanging system prompts.
   • Smart Prompt Structuring: Append dynamic information (e.g., current date) to the end of a prompt to maintain cacheability for the static beginning.
   • Demo Result: 90% lower input costs when cached.
3. Persistent Reasoning:
   • API exposes the ability to save reasoning turns across calls.
   • Benefits: Improves continuity of answers and cache efficiency.
4. Compaction:
   • Reduces input tokens for long user prompts or extensive tool calls by compacting conversation history.
   • Can be automatic or explicitly triggered.
   • Demo Result: 82% fewer input tokens for the same task, but with a latency hit for compaction.

Customer Spotlight: Ploy’s Migration to GPT-5.6

Bryant Chou (Co-founder & CTO of Webflow, now Founder of Ploy) and Lorenzo Gentile (AI Engineer, Ploy) share their experience. Ploy is an AI-powered marketing platform that turns websites into revenue-driving engines, offering SEO/AEO optimization, lead activation, and expert playbooks.

Ploy’s Value Maxing Strategies:

• Understanding KV Caching: Standard agent loops process N^2 tokens as context grows. Caching reduces cost by 10x for previously processed tokens.
• Avoiding Cache Breaks: Cache breaks if the context window isn’t append-only. Dynamic tool loading or updating system prompts (e.g., current time) at the beginning of the context will break the cache, leading to 10x cost.
• On-Demand Tool Loading without Breaking Cache: Ploy loads niche tool schemas to the end of the context window when needed, keeping core tools persistent.
  • Result: 45% reduction in tool schema tokens, 33% cost reduction in evals.
• Breakpoints for Cache Hit Rate:
  • Caching system prompts and core tools (which are consistent across chats).
  • Caching workspace memory (if it remains unchanged across multiple chats).
  • Result: 89% reduction in first message cost in a new chat, 5% overall token spend reduction in production.
• Batching Tool Calls:
  • Prompting the agent to combine independent tool calls into a single step reduces re-processing and “thinking” tokens.
  • Designing tools to perform multiple actions in one call saves output tokens.
  • Result: 14% cost reduction with the same pass rate and increased speed.
• General Tool Optimization:
  • Using web search features like Exa’s “highlights” to slim down responses: 70% reduction in web tool response size, estimated $37,000/year savings.
  • Making deterministic actions outside the agent’s LLM calls: 54% reduction in token spend.

Q&A Highlights

• Ploy’s Migration Rationale: Sought “alpha” and continually stress-tested new models like GPT-5.6 against real-world use cases (marketing strategy, page redesign, site import).
• GPT-5.6 Ultra Effort and Sub-agents:
  • Max: A new reasoning level above “Extra High” that fully unleashes the model’s reasoning capabilities.
  • Pro: A reasoning mode (now a parameter) that spends significantly more tokens and time for answers.
  • Ultra: A mode built on top of Max, modifying the system prompt to push for more sub-agent usage. It is extremely token-intensive and best for complex, heavy-hitting tasks.
  • Sub-agents are dispatched with only the needed context, not the full chat history, unlike side chats.
• Agent Depth and Performance:
  • Caution advised beyond two delegation layers due to context loss.
  • More effective for parallelizable work with independent lanes.
  • Performance heavily depends on the model’s orchestration abilities (e.g., Sol is better than older models).
  • Proceed with caution; older models often performed better without sub-agents in terms of cost/time.
• New Chat vs. Compaction:
  • Compaction is generally preferred for long threads, allowing continuous work in one thread for days/weeks.
  • A new chat might be better if the initial setup is heavily loaded with many plugins/skills/files that are always relevant and the model needs to reference them for every answer.
  • A UX problem exists: when should a user start a new chat for an entirely separate task?
• Downsides of Caching: Can become brittle if introduced without a good understanding of application parts that truly benefit, or if not carefully managed.
• Compaction Across Models/Reasoning Levels: The underlying endpoint mechanism works the same, converting conversation history into a reusable representation. The specific encoding and resulting token usage might vary slightly based on the model, reasoning level, and chat context.

This session highlights that maximizing value with AI requires a strategic approach to model selection, API feature utilization, and careful optimization of agent architecture and tool interactions.

• Tags:
• AI Agents
• Feature
• Evals
• AI Engineering
• Codex
• Cost Optimization
• Performance Optimization
• Prompt Caching
• API Best Practices
• Marketing Automation
• Webflow
• GPT-5.6
• Value Maxing
• Programmatic Tool Calling
• Context Compaction
• Ploy AI
• LLM Migration
• Reasoning Levels
• Tool Optimization