Choosing GPT-5.6 Sol, Terra, and Luna by Workload

Choosing a GPT-5.6 model does not require assigning the most expensive model to every request.

Start with the official model page: record model IDs, prices, context length, maximum output, and reasoning effort.

Then escalate only tasks with a high cost of failure and move verifiable routine work to a lower-cost model.

The values below were checked against official pages in August 2026.

Models, aliases, and prices can change, so recheck the official model list and model comparison during implementation.

Put the official values in one table

Item	Sol	Terra	Luna
API model ID	`gpt-5.6-sol`	`gpt-5.6-terra`	`gpt-5.6-luna`
Input / 1M tokens	$5.00	$2.00	$0.20
Cached input / 1M tokens	$0.50	$0.20	$0.02
Output / 1M tokens	$30.00	$12.00	$1.20
Context window	1,050,000	1,050,000	1,050,000
Maximum output	128,000	128,000	128,000
Knowledge cutoff	2026-02-16	2026-02-16	2026-02-16

The three models belong to the same GPT-5.6 family, but their prices and intended volume differ.

The model page lists none, low, medium, high, xhigh, and max as reasoning-effort options.

Reasoning effort is a separate axis from model selection.

Check the latest model guide for the current gpt-5.6 alias target and positioning.

Tasks to escalate to Sol

Choose Sol when the cost of failure is high, not merely because its unit price is highest.

Design involving constraints across several domains
Decisions that synthesize a long investigation
Changes with expensive review failures
Work where Terra’s representative success rate is insufficient

Sol does not need max reasoning effort for every request.

First compare Terra and Sol with the same effort to measure the model difference.

Then change only Sol’s effort and measure the incremental benefit.

Tasks for a Terra baseline

Terra is a practical baseline for a new workload.

Everyday implementation and code review
Understanding an existing codebase
Refactoring of moderate complexity
Work that needs a balance of quality and cost

Escalate only clearly failed Terra tasks to Sol.

Move tasks with excessive quality and easy verification down to Luna.

Defining escalation and downgrade criteria in advance makes model selection less subjective.

Conditions for moving work to Luna

The condition for Luna is not that a task looks easy.

The condition is that tests, schemas, rules, or retries can detect output errors.

Classification, extraction, and formatting
Small, clearly specified changes
First-pass summaries of large logs
Subagent work with automated error detection
Retryable work that does not add side effects on rerun

Do not choose Luna only by unit price for high-risk decisions whose errors cannot be detected.

Recalculate a pricing estimate

Assume that one operation uses one million input tokens and 100,000 output tokens.

Ignoring cache, tool charges, and long-context pricing, the estimate is:

Model	Input	Output	Total
Sol	$5.00	$3.00	$8.00
Terra	$2.00	$1.20	$3.20
Luna	$0.20	$0.12	$0.32

API usage records can separate regular input, cached input, and output tokens.

For an actual bill, map the response usage fields to the units in the official price table before adding them.

When one request exceeds 272K input tokens, check the long-context pricing on the official model page.

That condition applies a 2x input and 1.5x output multiplier to the full request.

If every request sends a huge context, revisit partitioning, caching, and removal of irrelevant history before changing models.

Compare with reasoning effort included

Keep the prompt, tools, timeout, and input material fixed during a model comparison.

Then keep the model fixed and change only reasoning effort.

Run representative tasks three to five times and record:

Success rate
Human correction count
Input and output token counts
Latency
Price
Failures detected by automated checks

The goal is not to crown the highest-tier model.

It is to find the lowest-cost combination that meets the required success rate for each task’s cost of failure.