GPT-5.6 API: Developer Guide to Using Sol, Terra & Luna

GPT-5.6 ships with three models and five reasoning levels — each with a different Model ID, price, and behavior. For developers, the question isn't "which is strongest?" but "which Model ID do I send in my API call for this task?"

This technical guide covers everything: Model IDs, pricing with examples, caching, Programmatic Tool Calling, and Ultra Mode.

Model IDs

Don't use gpt-5.6 if you want a specific model — this alias routes to Sol.

| Model | Model ID | Input Price | Output Price |

| ----- | --------------- | ----------- | ------------ |

| Sol | gpt-5.6-sol | $5/M | $30/M |

| Terra | gpt-5.6-terra | $2.50/M | $15/M |

| Luna | gpt-5.6-luna | $1/M | $6/M |

⚠️ The shorthand gpt-5.6 routes to Sol. Use the full ID for Terra or Luna.

Reasoning Levels

GPT-5.6 offers 5 reasoning levels controlling output token depth:

• medium (default) — standard reasoning
• high — deeper, for moderately complex tasks
• xhigh — deep reasoning for complex tasks
• max — deepest per-task reasoning
• ultra — 4 parallel agents working together (Sol only)

Higher levels = slower response + higher cost. For simple tasks, medium is sufficient.

Example API call with reasoning level

python
from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.6-sol",
    input="Analyze this data and give me a comprehensive report",
    reasoning_effort="high"
)

Programmatic Tool Calling

New in GPT-5.6 — the model filters intermediate data internally instead of sending everything back to you. This reduces round-trips and lowers costs.

Without Programmatic Tool Calling, every step sends all data back and forth.

With Programmatic Tool Calling:

• User → API: "Check weather and analyze"
• API → User: Final analysis only

Savings: 40-60% fewer tokens for multi-tool tasks.

Prompt Caching

GPT-5.6 supports advanced caching — the first request writes the cache, subsequent identical prefixes read from cache at 90% discount.

Cache write price: 1.25x input price
Cache read price: 0.1x input price (90% off)
Minimum cache life: 30 minutes

Example: A 10,000-token system prompt sent with every call — after the first request, 9,000 tokens become cached, and costs drop dramatically.

Ultra Mode — 4 Parallel Agents

ultra works on Sol only and launches 4 agents collaborating on the same task. Ideal for:

• Large codebase projects
• Research needing parallel exploration
• Very complex multi-step tasks

python
response = client.responses.create(
    model="gpt-5.6-sol",
    input="Build a complete API with database integration",
    reasoning_effort="ultra"
)

Practical Code Examples

1. Smart Routing

Start with Luna for simple tasks, Terra for moderate, Sol for complex:

python
def route_task(task_type, prompt):
    if task_type == "classification":
        model = "gpt-5.6-luna"
    elif task_type == "analysis":
        model = "gpt-5.6-terra"
    elif task_type == "complex_coding":
        model = "gpt-5.6-sol"

    return client.responses.create(model=model, input=prompt)

2. Cache for Repeated Queries

For support systems with repeated system prompts:

python
# First call: writes cache
response1 = client.responses.create(
    model="gpt-5.6-terra",
    input="You are a technical support agent..."  # System prompt + user question
)

# Second call: reads from cache (90% discount)
response2 = client.responses.create(
    model="gpt-5.6-terra",
    input="You are a technical support agent..."  # Same prefix → cache hit
)

3. Cost Estimator

Estimate monthly costs:

• Luna: $1/M input + $6/M output
• Terra: $2.50/M input + $15/M output
• Sol: $5/M input + $30/M output

Example: 200 requests/day, avg 4K input + 500 output tokens:

• Luna: 200 × ($1×4 + $6×0.5) = 200 × $7 = $1,400/month
• Terra: 200 × ($2.50×4 + $15×0.5) = 200 × $17.5 = $3,500/month
• Sol: 200 × ($5×4 + $30×0.5) = 200 × $35 = $7,000/month

With routing: Luna for 70%, Terra for 25%, Sol for 5% → ~$2,100/month instead of $7,000.

FAQ for Developers

Is GPT-5.6 API available in the UAE?

Yes, the UAE is included in the global rollout. All API endpoints work.

Does Luna support Function Calling?

Yes — Luna supports Function Calling, but with limited capability compared to Sol. Fine for simple tasks.

Best Model ID for Code Review?

Sol: gpt-5.6-sol with reasoning_effort="high". Terra: gpt-5.6-terra for small PRs.

Does GPT-5.6 support Structured Outputs?

Yes — JSON mode and Structured Outputs are supported across all three models.

📚 See the full comparison: Sol vs Terra vs Luna

📚 GPT-5.6 for UAE businesses

How Katbi Can Help

We build professional API integrations with GPT-5.6 — model selection, routing optimization, and cost management.

📱 Contact us on WhatsApp