After studying this material, you should be able to:

1. Identify the three tiers of the GPT-5.6 model family and their intended use cases
2. Explain the key capability improvements over previous models
3. Compare performance and cost trade-offs across model tiers
4. Understand the safety architecture and its layered approach
5. Apply knowledge of availability and pricing to select the right model for a given task

---

Step-by-Step Study Guide

Step 1: Understanding the Model Family Structure

Think of GPT-5.6 like a product lineup — similar to how phones come in Standard, Pro, and Budget versions.

GPT-5.6 Family
├── Sol        → Flagship (highest capability)
├── Terra      → Balanced (everyday work)
└── Luna       → Most affordable (cost-efficient)

Key Concept: Capability Tiers Are Durable

The number (5.6) identifies the generation Sol, Terra, Luna are permanent tier names that can improve independently

Think of it this way:

• Sol = A senior expert consultant
• Terra = A skilled professional
• Luna = A reliable generalist

---

Step 2: Understanding Performance Settings

GPT-5.6 introduces escalating effort levels — you choose how hard the model works:

Effort Levels (Low → High)
─────────────────────────────────────────
medium → xhigh → max → ultra
─────────────────────────────────────────
Speed   ←────────────────→  Power
Cost    ←────────────────→  Quality

What Each Level Does

Key Insight About ultra

Ultra trades higher token cost for faster time-to-result by running multiple agents simultaneously — like assigning a task to four specialists working in parallel instead of one

---

Step 3: Core Capability Improvements

3A: Coding

GPT-5.6 Sol is described as OpenAI's best coding model yet.

Benchmark Results (Artificial Analysis Coding Agent Index):

GPT-5.6 Sol (max)    ████████████████████ 80.0  ← New state of art
Claude Fable 5       ████████████████████ 77.2
GPT-5.5              ███████████████████  76.4
GPT-5.6 Terra        ███████████████████  77.4
GPT-5.6 Luna         ██████████████████   74.6

Efficiency Advantage:

• Uses less than half the output tokens vs. competitors
• Completes tasks in less than half the time
• Costs approximately one-third less

New Feature: Programmatic Tool Calling

This is an important new capability to understand:

Traditional Approach:
User → Model → Tool → Model → Tool → Model → Result
       (every step passes through the model)

Programmatic Tool Calling:
User → Model writes mini-program → Program runs tools autonomously → Result
       (model only involved at start and end)

Why this matters:

• Fewer model round trips = lower cost
• Filters large intermediate data automatically
• Adapts workflow without constant guidance

3B: Knowledge Work

GPT-5.6 integrates with real workplace tools:

Supported Integrations
├── Slack
├── Notion
├── Microsoft 365
└── Google Drive

Standout Results:

• BrowseComp: 92.2% (web research accuracy)
• OSWorld 2.0: 62.6% (computer use tasks) — using 85% fewer tokens than competitors

3C: Design Capabilities

This is a qualitative leap worth understanding:

GPT-5.5 Approach:          GPT-5.6 Approach:
Generate code/content  →   Generate + Inspect rendered result
                           + Catch visual issues
                           + Apply finishing touches
                           + Deliver ready-to-use output

Practical Example — Presentation Templates:

GPT-5.6 can read a deck's design system (fonts, colors, spacing, Slide Master rules) and apply those conventions consistently to new content — something GPT-5.5 failed to do reliably

3D: Cybersecurity

Important Dual-Use Concept:

Same Capability
     │
     ├──→ Offensive: Find vulnerabilities to exploit
     └──→ Defensive: Find vulnerabilities to patch

Benchmark Improvements:

Trusted Access Program:

Verified cybersecurity professionals can unlock additional defensive capabilities through OpenAI Daybreak's Trusted Access for Cyber program

3E: Science

GPT-5.6 shows improvements across life sciences:

---

Step 4: Understanding the Safety Architecture

This is one of the most conceptually important sections.

The Layered Safety Model

Safety Stack (Multiple Layers)
┌─────────────────────────────────────┐
│  Layer 1: Training-time protections │  ← Built into model weights
├─────────────────────────────────────┤
│  Layer 2: Real-time checks          │  ← Active during inference
├─────────────────────────────────────┤
│  Layer 3: Reasoning monitor         │  ← Reviews full conversation context
├─────────────────────────────────────┤
│  Layer 4: Continuous monitoring     │  ← Post-deployment surveillance
├─────────────────────────────────────┤
│  Layer 5: Account-level enforcement │  ← Trust-based access control
└─────────────────────────────────────┘

Key Innovation: Reasoning Monitor

Traditional safety systems use classifiers — simpler models that flag keywords or patterns.

GPT-5.6 adds a reasoning monitor that:

• Reviews the full conversation context
• Determines intent and likely consequences
• Can be updated rapidly without retraining classifiers

Old Approach:          New Approach:
Keyword/Pattern   →    Context-aware reasoning
  flagging              about harm potential
  (binary)              (nuanced)

Safety Thresholds

OpenAI uses a Critical threshold concept:

Risk Scale
──────────────────────────────────────
Low Risk → Medium Risk → HIGH RISK → CRITICAL
                                         ↑
                              GPT-5.6 does NOT
                              cross this line in
                              biology or cybersecurity

The Overblocking Problem

Key insight: Being too restrictive creates its own security risk

Overblocking Consequences:
├── Defenders cannot test their own systems
├── Defenders cannot deploy patches
└── Malicious actors use other (open-source) models anyway

Result: Net harm to security, not net benefit

---

Step 5: Pricing and Availability

Pricing Structure

Model    Input (per 1M tokens)    Output (per 1M tokens)
─────────────────────────────────────────────────────────
Sol      $5.00                    $30.00
Terra    $2.50                    $15.00
Luna     $1.00                    $6.00

Cost Ratio Visualization:

Sol    ██████████████████████████████  $30 output
Terra  ███████████████                 $15 output
Luna   ██████                          $6 output

New Caching System

Cache Behavior:
├── Cache writes:  1.25x normal input rate
├── Cache reads:   90% discount on input
└── Minimum cache life: 30 minutes

Practical meaning: If you repeatedly send similar long prompts (like a system prompt), caching saves significant cost after the first call.

Access by User Tier

---

Step 6: Synthesizing the Core Value Proposition

Now that you understand the components, here is the central thesis of GPT-5.6:

The Core Trade-off Problem in AI Models:
─────────────────────────────────────────
More capability = More tokens = Higher cost = Slower

GPT-5.6's Claimed Solution:
─────────────────────────────────────────
More capability + Fewer tokens + Lower cost + Faster
         ↑
"More useful work from every token"

Evidence Supporting This Claim

---

Quick Reference Summary

┌─────────────────────────────────────────────────────┐
│                  GPT-5.6 At a Glance                │
├─────────────────────────────────────────────────────┤
│ MODELS:   Sol (flagship) > Terra (balanced)         │
│           > Luna (affordable)                       │
├─────────────────────────────────────────────────────┤
│ SETTINGS: medium → xhigh → max → ultra              │
│           (efficiency ←→ power)                     │
├─────────────────────────────────────────────────────┤
│ KEY WINS: Coding, Computer Use, Cybersecurity,      │
│           Science, Knowledge Work, Design           │
├─────────────────────────────────────────────────────┤
│ SAFETY:   Layered system, reasoning monitor,        │
│           Trusted Access for sensitive domains      │
├─────────────────────────────────────────────────────┤
│ PRICING:  Sol $5/$30 | Terra $2.50/$15 | Luna $1/$6 │
│           (input/output per 1M tokens)              │
└─────────────────────────────────────────────────────┘

---

Self-Check Questions

Test your understanding:

1. What is the difference between Sol, Terra, and Luna? (Hint: capability tiers, not just sizes)
2. Why does ultra cost more but potentially save money? (Hint: think about time-to-result)
3. Why is overblocking a cybersecurity risk? (Hint: who still has access?)
4. What makes Programmatic Tool Calling more efficient? (Hint: model round trips)
5. A startup needs strong coding performance on a tight budget. Which model and why? (Hint: compare Terra's benchmark scores vs. cost)