OpenAI GPT-5.6 Luna 80% price cut!

 GPT-5.6 just made itself better... - YouTube by  Matthew Berman

OpenAI released a major model update focused on recursive infrastructure self-improvement, knowledge distillation, and tier-based model routing. Instead of relying solely on human engineers to optimize their serving stack, OpenAI used its flagship model, GPT-5.6 Sol, to rewrite and refactor its own low-level backend kernels (including CUDA and serving code).

By lowering infrastructure overhead through automated kernel optimization, OpenAI drastically reduced serving costs across their entire stack. This enabled significant price cuts and rate-limit increases, positioning OpenAI aggressively against low-cost competitors like Chinese model Kimi K3.

1. Official Pricing & Model Family Breakdown

• GPT-5.6 Luna (Volume / Lightweight Worker)

  • Price Drop: Cut by 80%

  • New Pricing: $0.20 per 1M input tokens / $1.20 per 1M output tokens (down from $1.00 / $6.00)

  • Use Case: High-speed grunt work, background code execution, automated code reviews, and sub-agents.

• GPT-5.6 Terra (Balanced / Standard Workloads)

  • Price Drop: Cut by 20%

  • New Pricing: $2.00 per 1M input tokens / $12.00 per 1M output tokens (down from $2.50 / $15.00)

  • Use Case: Everyday general tasks, standard coding, and balanced reasoning.

• GPT-5.6 Sol (Flagship / Heavy Reasoning & Planner)

  • Pricing: Retained at $5.00 per 1M input tokens / $30.00 per 1M output tokens.

  • New Feature: Fast Mode offering up to 2.5x execution speed at a 2x price multiplier.

  • Use Case: High-level architecture, complex orchestration, task planning, and generating distilled knowledge for smaller worker models.

---

2. Technical & Workflow Innovations

• Recursive Infrastructure Optimization: Rather than generating user-facing application code, Sol was pointed inward at OpenAI's production kernels—achieving real-world recursive self-improvement at the systems layer.

• Planner/Worker Orchestration: Developers using environments like OpenAI Codex are shifting away from using single large models for entire tasks. Instead, Sol acts as a planner that dispatches tasks to low-cost Luna worker instances.

• Knowledge Distillation: Knowledge and capabilities from flagship models like Sol are actively distilled into the Luna and Luna Max models, preserving high code accuracy at a fraction of the compute cost.

---

3. Strategic & Industry Shift

• The "Kimi K3 Effect": Low-cost, high-performance models from competitors like Moonshot AI (Kimi K3/K4) have forced Western AI providers to lower costs and boost throughput to stay competitive.

• Cost Per Task vs. Benchmarks: Evaluators and developers are moving away from raw benchmark chasing ("benchmaxxing") toward total cost per completed task and task completed correctly without silent regressions.