Every new model release triggers the same tired debate. Commentators rush to crown a winner and declare the previous tier dead. With GPT-5.6 Luna sitting alongside Terra and Sol, the narrative writes itself: Luna is cheap enough and capable enough to make Terra irrelevant. This is wrong. It is also expensive. Picking a model for your coding agents is not a beauty contest, a team identity, or a benchmark horse race. It is an operating policy. The teams that internalize this distinction will spend less, move faster, and break fewer things than the teams that default to the strongest model for every request.
Your Default Should Be the Cheapest Tool That Fits
Luna is the value tier, and that is not faint praise. It thrives on bounded, explicit, and easy tasks. Think classification, summarization, short code edits, and first-pass research. When an agent parses a support ticket to assign a priority label, Luna is enough. When it renames a variable across a few files or drafts a one-paragraph summary of a git diff, Luna is enough. These are jobs with narrow scopes, clear inputs, and objectively checkable outputs.
The economic effect is what changes the game. Luna is cheap. At high volume, this shifts automation from an expensive ceremony to infrastructure. You stop counting tokens and start measuring throughput. A cheap model that clears eighty percent of routine tasks is more valuable than an expensive model that clears eighty-five percent if that extra five percent does not change the outcome. If Luna generates a unit test in two seconds and Terra generates a marginally cleaner one in eight seconds for five times the cost, the math only works if someone is carefully auditing every line. Most of the time, no one is. Luna should be your default for bounded work precisely because most work is bounded.
Escalate When the Boundaries Disappear
Terra is not useless. It is your escalation tier, and it earns its keep on tasks without clear boundaries. Use it when the goal is underspecified or when the work involves complex systems like deployment paths, cross-module changes, or incident triage. A deployment path that snakes through staging, canary, and production environments with feature flags does not have a tidy spec sheet. A refactor that touches the billing logic and quietly ripples into the reporting pipeline is not a bounded task. A production incident where the logs scream about API timeouts but the root cause lives in a migration script from last quarter requires judgment.
Terra provides that judgment. It separates symptoms from causes. Luna might patch a retry loop to stop the bleeding. Terra asks whether the retry loop should exist at all, or whether the underlying timeout architecture is the real problem. That difference matters when the wrong fix turns a temporary slowdown into a cascading failure. A stronger model that prevents one bad production migration is worth the price if it saves an engineer a full day of cleanup. One prevented outage pays for months of escalation margin.
Sol Is the Insurance Policy, Not the Daily Driver
Sol exists for cases where extra capability justifies the high cost. Use it for high-risk reviews or architectural changes. Rebuilding the authentication flow, redesigning database sharding, or approving a pull request that touches the payment gateway are not daily occurrences. They are events. Sol should not be your default. It should be your exception handler, summoned when the cost of failure is too high for cheaper models to carry alone.
For the highest-risk category, pair Sol with a deterministic verifier. Let Sol suggest the schema change or reason through the architectural trade-offs. Let your CI pipeline, static analysis, and integration tests confirm the mechanical details. The model brings intuition. The verifier brings guarantees. That combination is what protects you when the blast radius is largest.
Build a Router, Not a Religion
The real metric is not which model is best. The question is which model should handle this task based on cost, latency, and blast radius. Stop treating model choice as an identity. Do not say, "We are a Terra shop." Instead, route by task class.
간단한 분류기를 구축하세요. 들어오는 작업은 영향 범위(blast radius)에 따라 태그가 지정됩니다. 영향 범위가 낮은 작업은 Luna로 보냅니다. 중간 정도의 영향 범위를 가진 작업은 Terra로 보냅니다. 영향 범위가 큰 작업은 강력한 모델과 결정론적 검증기(deterministic verifier)로 보냅니다. 시작할 때 완벽한 머신러닝 분류기가 필요하지는 않습니다. 몇 가지 휴리스틱(heuristics)만으로도 충분합니다. 내부 유틸리티만 건드리고 코드 라인 수도 적은 코드 리뷰인가요? Luna입니다. 배포 파이프라인, 서비스 간 호출, 또는 모호한 요구사항을 언급하는 티켓인가요? Terra입니다. 고객 데이터, 핵심 경로(critical paths), 또는 법적 준수 사항을 건드리는 모든 것인가요? Sol로 에스컬레이션하고 사람 또는 결정론적 검토를 거치도록 하세요.
모델 이름이 아니라 결과를 측정하세요. 작업당 비용, 재시도율, 그리고 누락된 결함(escape defects)을 추적하세요. Luna에 할당한 작업에서 실패가 발생한다면, 경계선을 위로 올리세요. 매일 반복되는 패턴에 Terra가 과하다면, Luna로 강등시키고 비용(burn)이 줄어드는 것을 확인하세요. 목표는 예산을 초과하지 않으면서 자동화를 늘리는 것입니다. Luna는 대량의 백그라운드 작업을 처리합니다. Terra는 판단이 중요한 순간을 처리합니다. Sol은 한 주를 망칠 수 있는 예외 상황을 감시합니다.
이를 제대로 수행하는 팀은 에이전트 군단(agent fleet)을 잘 운영되는 엔지니어링 조직처럼 다룹니다. 모든 프로젝트에 아키텍트를 배치하지 않으며, 인턴에게 핵심 데이터 모델을 재설계하라고 요구하지도 않습니다. 그들은 역량을 리스크에 맞게 매칭합니다. 여러분의 모델에도 똑같이 적용하세요.
원문 토론 읽기: GPT-5.6 Luna Is The Value Tier. Terra Is Not Useless
GyaanSetu 학습 커뮤니티 참여하기: t.me/GyaanSetuAi