GPT-5.6 Terra and Luna: Balance and Volume

Status as of November 5, 2026. The table below reflects current pricing: since July 30, 2026, Luna costs 80% less and Terra 20% less than at launch, when they were billed at $1 / $6 and $2.50 / $15 respectively. For new projects, the GPT-6 family is even cheaper: GPT-6 Luna at $0.10 / $0.50, GPT-6.1 Terra at $1 / $5, and GPT-6.1 Sol at $2 / $10—barely more than GPT-5.6 Terra (Lesson 9). GPT-5.6 Terra and Luna remain the official replacements for gpt-5-mini and gpt-5-nano, which were retired on December 11.

The lighthouse isn’t for every job. Paying Sol’s rate to classify support tickets or extract dates from invoices is wasteful. That’s exactly why the GPT-5.6 family comes in three models. OpenAI positions Terra as the balanced choice for daily work and Luna as its fastest, most economical option. The right pick doesn’t depend on a prestige hierarchy but on the quality-to-cost ratio, task by task.

The three models side by side (as of November 5, 2026)

Three billing rules complete this picture. Beyond 272,000 input tokens, the entire request shifts to the “long context” rate: double for input, 1.5× for output. That means $8 / $30 for Sol, $4 / $18 for Terra, and $0.40 / $1.80 for Luna. Batch processing and Flex processing cost half price. Finally, since GPT-5.6, writing to the 30-minute cache is billed at 1.25× the input price, and 2.5× for the 12-hour cache introduced on October 20; reading from the cache keeps its 90% discount.

The last row of the table is identical across all three columns. All three models accept text and images as input, offer the same 1.05-million-token context window, and support the same reasoning effort levels, from none to max. The difference lies in the capacity-to-price ratio, not the feature list.

How to choose

Terra is a solid starting point for daily work: writing, summarizing, assistance, and routine coding. At $2 input and $12 output, it costs half as much as Sol on input and 40% less on output. A simple rule: start with Terra, and only pay Sol’s rate when Terra has concretely let you down on your own cases.

Sol keeps its place when quality comes first: difficult code, long-running agents, and analysis where an error is costly. The price gap is justified here because human rework—reviewing, correcting, and restarting—costs far more than the few dollars saved on the API call.

Luna targets high volume. At $0.20 input and $1.20 output per million tokens, it’s built for classification, extraction, sorting, and massive processing jobs where each request is simple but you’re running tens of thousands of them. On these workloads, the price gap with Sol—a factor of 17 to 20—shows up directly on your bill.

The cascade strategy

The classic pattern works well for this family: process the main flow with Luna, escalate to Terra what fails or needs nuance, and reserve Sol for cases that justify its price. Since all three models share the same API, inputs, and effort settings, switching between them boils down to changing the identifier.

The skeleton of a cascade router fits in a few lines:

def choose_model(task):
    if task.complexity == "high":
        return "gpt-5.6-sol"
    if task.complexity == "medium":
        return "gpt-5.6-terra"
    return "gpt-5.6-luna"

In a real system, the complexity criterion could be a business heuristic (document length, request type), a score produced by Luna itself, or simply the failure of a first attempt. The principle remains the same: every request is handled by the cheapest model capable of doing the job. The same logic applies to the GPT-6 family, with its own benchmarks: OpenAI’s documentation recommends GPT-6 Luna for volume, GPT-6.1 Sol for balance, and GPT-6 Astra for the most demanding tasks.

Your turn

Price a cascade. Take a flow of 100,000 requests per month, each with 2,000 input tokens and 500 output tokens. Calculate its cost using Sol alone, Terra alone, Luna alone, and then a cascade that sends 80% of requests to Luna, 15% to Terra, and 5% to Sol. You should find approximately $1,800, $1,000, $100, and $320. This calculation ignores reasoning tokens, which are billed as output tokens: redo it with your actual volumes and the effort level you plan to use.

Official sources: GPT-5.6 Terra Spec Sheet · GPT-5.6 Luna Spec Sheet · Pricing · Models (OpenAI API documentation) · July 30 Price Cut Announcement (OpenAI)

Key takeaways

• Terra ($2 / $12) is the balance for everyday use; Luna ($0.20 / $1.20) targets massive workloads
• These prices include the July 30 reduction; long context applies beyond 272,000 tokens; Batch and Flex are half price
• All three models share the same API, inputs, context window, and effort levels: only the identifier changes
• The Luna, Terra, Sol cascade cuts costs without sacrificing difficult cases; the GPT-6 family does better for less