Choice Request | GPT-5.6 Luna, Terra, or Sol? Price-aware routing backed by one local comparison

Direct answer

Use GPT-5.6 Luna first for clearly bounded execution that is cheap to verify, Terra for routine multi-step development, and Sol at planning, review, or decision points where failure is expensive. This is not a universal ranking. It is a routing rule: define the acceptance line, then choose the least expensive model that can cross it reliably.

This recommendation combines two different kinds of evidence. OpenAI's July 30, 2026 announcement establishes API prices. Token Beggars LAB-0001 only establishes which correctness boundaries were observed in one fixed local coding task.

GPT-5.6 API prices

OpenAI reported an 80% price reduction for Luna and 20% for Terra:

• Luna: $0.20 per million input tokens and ".20 per million output tokens.
• Terra: $2 per million input tokens and "2 per million output tokens.
• Sol: unchanged standard pricing; Fast mode can reach up to 2.5 times standard speed at twice the standard price.

Luna and Terra differ by an order of magnitude on listed input and output prices. That does not make token price the full cost. A failed run may trigger rework, human review, rollback, data repair, or a production incident. The useful metric is total cost per accepted result.

Route by failure cost

Start with Luna when acceptance is cheap

Good candidates have explicit inputs and outputs plus inexpensive automated checks: mechanical file edits, unit-test generation, structured data cleanup, fixed-format drafts, or a scoped component that can be judged by lint, typecheck, and tests. Luna's price advantage matters only while the result stays above the acceptance line.

Use Terra for routine multi-step development

Terra is a practical default when the task requires code search, edits in several places, tool calls, and iteration, but remains reversible and testable. Retry behavior, concurrency, authorization, account isolation, and persistence still need to be written into acceptance criteria; “the happy path works” is not enough.

Put Sol at high-cost boundaries

Sol is better placed on planning, risk enumeration, architecture review, and final inspection when a wrong decision could corrupt data, expose account state, or create irreversible effects. Typical boundaries include authentication recovery, idempotency, database migrations, authorization, concurrency, and release review. A cost-aware combination is: Sol defines the boundaries, Luna or Terra implements the bounded work, automated tests enforce them, and Sol reviews the critical diff.

What LAB-0001 showed

LAB-0001 used one preregistered rescue task: carry an anonymous draft through login, submit the same request at most once, and never move Account A's draft into Account B. All candidates used the same repository snapshot, task, and time limit without extra prompts during the run.

All three passed the ordinary local browser path. Sol and Luna also passed same-account server replay and cross-account isolation checks. Terra passed the UI path, but a later local audit cleared cooldown state to simulate a retry after cooldown; the same authenticated session and request body then produced two 201 responses with different Post IDs.

Sol and Luna therefore tied on this task's correctness-first score. The reusable lesson is not the rank: a browser flow passing does not prove server idempotency, and the tested failure boundary can change whether a cheaper model is truly cheaper.

The experiment did not wait five real minutes, simulate transport-level packet loss, or run in production. Formal token totals for Sol and Luna and actual monetary cost for all three candidates were not recorded. It cannot support a general model leaderboard or cost ranking.

A reusable routing checklist

1. Define the accepted outcome and required tests.
2. List boundaries that cannot be lost: authorization, concurrency, idempotency, migration, privacy, or production data.
3. Estimate failure cost, not only token price.
4. Start with Luna or Terra on verifiable work and escalate only when the same boundary keeps failing.
5. Split the task: reserve Sol for planning, critical judgment, and final review instead of paying the highest rate for mechanical execution.

— Token Beggars Editorial

References

• Token Beggars LAB-0001