GPT Sol or GPT Terra: which one is cheaper?

By LW Forge — maintainer of LLM Scout · Updated August 31, 2026

GPT-5.6 Terra is cheaper than GPT-5.6 Sol. At the standard short-context rates checked on August 31, 2026, Terra costs $2.00 per million input tokens and $12.00 per million output tokens. Sol costs $4.00 and $20.00. That means Terra is half the price for new input and 40% cheaper for output. The final decision still depends on your traffic, how much context can be reused, and whether Sol's extra capability prevents retries and human review.

This article answers the practical comparison behind the question and turns per-token prices into three workload calculations. For a forecast based on your actual traffic, use the OpenAI cost calculator. A generic example can show the direction, but input, output, and cached tokens vary dramatically from one application to another.

Quick price comparison

Both models belong to the GPT-5.6 family, but they occupy different positions. OpenAI describes Sol as the family's highest-capability model and Terra as the option that balances intelligence and cost. Terra is not a free tier or a fixed monthly plan: the API bills the tokens processed by each request.

Prices are per million tokens and were checked against OpenAI's official API pricing page on 2026-08-31. Requests with more than 272K input tokens use a long-context tier with different rates: Sol becomes $8.00/$30.00 and Terra becomes $4.00/$18.00 per million input/output tokens. Terra therefore remains cheaper in both regimes, although the ratio is no longer exactly the same.

Caching does not make Sol the cheaper option. A cached-input read costs $0.40 per million tokens on Sol and $0.20 on Terra. If an application repeats a long system prompt, both models get the same relative 90% read discount; because Terra starts with the lower rate, its absolute price remains lower. Cache writes have a separate rate and should only be included when the cached prefix is created or refreshed.

What the difference looks like in practice

The cleanest comparison is to hold the workload constant, calculate input and output separately, and then compare totals. The formula is straightforward:

cost = (input tokens / 1,000,000 × input rate) + (output tokens / 1,000,000 × output rate)

These examples exclude tool calls, priority processing, regional endpoints, and taxes. Those items can change the final invoice and should be checked in the provider's documentation. The goal here is to isolate the difference between Sol and Terra.

Scenario 1: customer-support chatbot

Imagine 10,000 conversations per month. Each conversation sends 2,000 tokens of history, instructions, and retrieved context and receives a 300-token answer. That is 20 million input tokens and 3 million output tokens.

Terra saves $64 per month, or about 46% of the Sol cost. This is the kind of workload where Terra should be the first model tested: support answers, intent classification, information retrieval, and reasonably predictable instructions. If an evaluation shows that Sol prevents costly mistakes, calculate cost per resolved conversation rather than cost per token alone. The guide to building a chatbot with GPT-5 helps separate model spend from the other components of a product; RAG cost per query shows why retrieved context and answer length matter so much to this type of bill.

Scenario 2: internal assistant

Now consider 1,000 internal-assistant tasks per month. Each task sends 20,000 tokens of documents and instructions and receives a 1,000-token answer. The monthly load is 20 million input tokens and 1 million output tokens.

Terra saves $48 per month. Sol may be justified for document synthesis with difficult constraints, multi-step planning, or tasks where an error carries a high cost. Even then, the decision should come from a test with real documents. Many applications choose Sol by reputation when Terra already produces a correct, structured result.

If the same material is sent repeatedly, evaluate prompt caching. With 90% of the 20M input tokens cached, only 2M would be new and 18M would be cached. That would reduce monthly input to $15.20 for Sol and $7.60 for Terra, before output. Terra still wins; caching simply reduces the repeated portion of the bill.

Scenario 3: complex coding and reasoning

For a coding agent, suppose there are 500 tasks per month, each with 30,000 input tokens across instructions, files, and history, and 3,000 output tokens. The monthly load is 15M input tokens and 1.5M output tokens.

Terra saves $42 on raw token cost. This is also the scenario where Sol has the best chance of earning its premium: a wrong patch can cause a production failure, another review, or several correction calls. The correct comparison is cost per accepted task. Run the same tasks on both models, fix the output cap, validate tests automatically, and charge retries to the model that caused them.

When should you choose each model?

Choose Terra as the default when you need predictable cost, higher volume, straightforward success criteria, and tasks that do not require maximum capability on every call. It is the sensible starting point for support, extraction, classification, structured generation, and internal assistants whose prompts are demanding but repeatable.

Choose Sol when the work needs more planning, difficult code, tool use, conflicting instructions, or a response whose human review costs more than the price difference. Even then, do not assume that “more expensive” means “cheaper per task.” Measure quality, retries, output tokens, and execution time.

You can also combine the two. A router can send most tasks to Terra and escalate only ambiguous cases to Sol. This is often more efficient than using one model for the whole application, provided the classifier and escalation rules are evaluated too. For a wider view of cheaper tiers, see the comparison of the cheapest LLM APIs.

How to run your own comparison

Start by measuring a real request sample. Separate new tokens, cached tokens, and output tokens; record rejected answers, retries, and tool calls. Then run the same task set on both models and calculate four numbers: cost per request, cost per correct task, retry rate, and projected monthly cost. The token calculator helps estimate prompt size, while the OpenAI cost calculator lets you test traffic assumptions.

Prices change, and an API list price is not necessarily the final price in your country. The LLMScout sources and methodology page explains how the catalog treats pricing and reference links. For this comparison, the authority is OpenAI's official pricing documentation; the date above is a snapshot, not a promise that the rate will remain unchanged.

Frequently asked questions

Which is cheaper: GPT Sol or GPT Terra?

GPT-5.6 Terra is cheaper. At the standard rates checked on August 31, 2026, it costs $2.00 per million input tokens and $12.00 per million output tokens, compared with $4.00 and $20.00 for Sol.

Does GPT Terra cost half as much as GPT Sol?

For new input, yes: $2.00 versus $4.00 per million tokens. For output, Terra costs $12.00 versus Sol's $20.00, so it is 40% cheaper. Your total depends on the mix of input, output, and cached tokens.

Is GPT Sol worth the extra cost?

It can be worth it when higher capability reduces errors, retries, or human review on complex tasks. A price table cannot guarantee that outcome; test both models with a representative evaluation and calculate cost per accepted task.

Does caching make Sol cheaper than Terra?

No. Cached-input reads cost $0.40 per million tokens on Sol and $0.20 on Terra. Caching lowers the cost of repeated prompt content for both models, but Terra still has the lower absolute rate in this comparison.

Prices and context rules were checked against OpenAI's official API pricing documentation and the GPT-5.6 Terra model page on 2026-08-31. Rates can change; confirm the source before finalizing a budget.