GPT-5.6 Sol, Terra, and Luna: Routing Policy for OpenAI's New Three-Tier Model Family

OpenAI began a limited preview of GPT-5.6 on June 26, 2026 — not as a single model but as a three-tier family: Sol (flagship), Terra (balanced), and Luna (fast and affordable). The preview is restricted to roughly 20 trusted organizations whose participation was coordinated with the US government, with general availability promised "in the coming weeks."

The routing implication is immediate even before the models reach GA: engineering teams now have three new model IDs to plan routing policies around, a new pricing structure that changes the cost case for the mid-tier, and two new reasoning modes — max and ultra — that alter how cost-per-task is calculated for agentic work.

What happened

OpenAI introduced three models under the GPT-5.6 umbrella:

• GPT-5.6 Sol — the flagship tier. Same price as GPT-5.5 ($5 / $30 per million tokens input/output), 1.5M-token context window (up from 1.05M in GPT-5.5 Pro), top-of-family on Terminal-Bench 2.1 for agentic coding. Sol runs in three modes: standard, max (deeper single-model reasoning), and ultra (parallel subagent orchestration).
• GPT-5.6 Terra — the balanced tier. $2.50 / $15 per million tokens — half the price of GPT-5.5 for equivalent capability according to OpenAI's stated positioning. This is the routing story: Terra is pitched as the new cost-efficient default for tasks that previously defaulted to GPT-5.5.
• GPT-5.6 Luna — the fast and affordable tier. $1 / $6 per million tokens, positioned as the replacement for high-volume, low-complexity workloads.

OpenAI has not published exact API model ID strings in the preview materials. Based on the naming convention of GPT-5.5, expect gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna once the API opens broadly.

The limited preview is API-only (no ChatGPT web access yet) and restricted to a small set of organizations. Broader availability is gated behind a US government review process — OpenAI acknowledged this is an "unsustainable long-term default" and has committed to a more systematic framework.

Why it matters for AI engineering teams

The most operationally significant number is Terra's price: $2.50 / $15 per million tokens. If Terra delivers GPT-5.5-class output quality on your workloads, that is a direct 50% input cost reduction relative to GPT-5.5 ($5 / $30). For teams already routing to GPT-5.5 as their default reasoning tier, Terra's GA arrival is a scheduled cost-reduction event — not a nice-to-have, but something to route-test in advance.

Luna at $1 / $6 represents the new lowest-cost OpenAI model available through the standard API, likely competing with GPT-4.1-nano or GPT-5.4-mini for high-throughput classification, extraction, and summarization workloads.

The context window expansion for Sol (1.5M tokens) matters for teams running long-horizon coding agents, legal document review, or large-codebase refactors where previous context limits forced chunking or multi-turn workarounds.

Sol's ultra mode — which orchestrates parallel subagents rather than a single model invocation — changes the cost accounting model for agentic tasks. A single ultra-mode call can fan out into multiple subagent completions, each billed separately. Teams using Codex or custom coding-agent pipelines need to evaluate whether ultra's throughput gain justifies the token fan-out cost.

The router/operator angle

The shift from a single flagship model to a three-tier family with durable named tiers (Sol, Terra, Luna will advance independently) is a structural change in how routing policies should be built against OpenAI's lineup.

Three routing lanes to plan now:

1. Sol lane — highest-capability tasks: long-horizon agentic coding, security research, biology/genomics analysis, multi-turn complex reasoning. At $5 / $30, cost profile is unchanged from GPT-5.5. If your current routing targets gpt-5.5 for flagship work, gpt-5.6-sol is the natural successor.

2. Terra lane — the new cost-efficient standard: everyday coding assistance, document summarization, multi-step reasoning for business logic, agentic tasks that don't require frontier capability. At 50% of GPT-5.5's price for equivalent stated capability, Terra is the most commercially interesting tier for most production workloads. Plan a routing rule that directs non-frontier tasks to gpt-5.6-terra at GA.

3. Luna lane — high-volume, low-complexity: classification, extraction, short summarization, intent parsing, structured output generation. At $1 / $6, Luna slots below Terra as a fallback for volume-sensitive workloads where quality requirements permit.

Fallback chain during limited preview: Until general availability, gpt-5.6-* model IDs will return errors for most API keys. Your routing configuration should treat 404/model-not-found responses on GPT-5.6 IDs as a GA-readiness signal, not a permanent error. For now, keep GPT-5.5 as the active default in production routing; add GPT-5.6 tiers to staging configuration and route-test against Terra as soon as your organization gains access.

Ultra mode and subagent cost attribution: If you are routing Codex or any coding agent that will run in ultra mode, ensure your usage-accounting pipeline can attribute fan-out completions back to the parent task. A single ultra-mode request is not a single completion event — it generates multiple subagent tokens under a shared parent. Check whether your cost-tracking middleware collapses these to a single row or handles child-completion attribution correctly.

Alias policy: Based on GPT-5.5 precedent, expect gpt-5.6 (without tier suffix) to resolve to Sol. Do not rely on the bare alias for cost-sensitive production workloads — pin to gpt-5.6-terra or gpt-5.6-luna explicitly once the API opens.

What to watch

• General availability announcement. GA is "coming weeks" from June 26 — watch the OpenAI API changelog for the exact model ID strings and pricing confirmation.
• Terra quality validation on your workloads. OpenAI's positioning of Terra as GPT-5.5-equivalent is a stated claim, not a benchmark-verified guarantee for your specific task distribution. Route 5–10% of staging traffic to Terra as soon as access opens and measure output quality against your current GPT-5.5 baseline before committing to full migration.
• Luna vs. existing mini-tier models. At $1 / $6, Luna enters a crowded field alongside GPT-4.1-nano and GPT-5.4-mini. Benchmark Luna against your current mini-tier model on your high-volume classification or extraction workload before treating it as an automatic replacement.
• Sol Ultra billing behavior. Confirm with OpenAI whether ultra-mode completions are billed as a single Sol completion or as separate subagent completions before enabling ultra mode in production.

For TheRouter users, routing rules that currently target gpt-5.5 by name will continue to work through the transition. Once GPT-5.6 IDs are confirmed live in the API, update your provider model list and define tier-specific routing rules — Terra for the cost-optimized lane, Sol for the capability lane.