Sol, Terra or Luna?
The GPT-5.6 tier selection guide
GPT-5.6 ships in three tiers: flagship Sol, balanced Terra and fast Luna. All three share the 1.05M API context and 128K output — they differ in capability, speed and price. This guide uses verified numbers to help you pick a default.
The three tiers, side by side
Figures from OpenAI's official release and docs (verified 2026-07-10).
| Metric | Sol | Terra | Luna |
|---|---|---|---|
| Positioning | Flagship — for the hardest problems | Balanced everyday tier — about half the price | Fastest and cheapest |
| Terminal-Bench 2.1 | 88.8% | 84.3% | 82.5% |
| Price (input / output, per 1M tokens) | $5 / $30 | $2.50 / $15 | $1 / $6 |
| Cached reads (−90%) | $0.50 | $0.25 | $0.10 |
| API context / max output | 1.05M / 128K | 1.05M / 128K | 1.05M / 128K |
| Speed | Up to ~750 tokens/s on Cerebras from July | Standard serving speed | Fastest of the three |
Note: Sol Ultra is Sol's high-compute mode (91.9% on Terminal-Bench 2.1), not a separate price tier.
Pick a tier by scenario
Choose Sol: the hardest engineering and reasoning work
Cross-repo refactors, long-horizon agent runs, hard debugging and architecture calls. Its 88.8% on Terminal-Bench 2.1 is the top score among GA models; on the hardest 10% of tasks the success-rate gap outweighs the price gap.
Choose Terra: the everyday default
Above-GPT-5.5 quality at about half of Sol's price (Terminal-Bench 84.3% vs GPT-5.5's 83.4%). For day-to-day feature work, code review and test generation, Terra is the best value default.
Choose Luna: high-volume, low-latency, batch
Autocomplete, commit messages, bulk classification and lightweight CI checks. At $1/$6 with the fastest responses, high-frequency calls stay cheap — and 82.5% on Terminal-Bench is unmatched at this price.
Cost examples
A typical agentic coding session (800K input + 60K output tokens): about $2.90 on Terra, $5.80 on Sol, $1.16 on Luna. With 80% of input hitting cache (reads at 10%), Terra drops to about $1.46. The tier choice moves cost by 2–5x for the same workload.
Switch per request — no commitment
The tiers are just three model ids: gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna. Switch any time with /model in Codex CLI, or per request in the API. Via QCode all three share one key and one quota — Sol for the hard parts, Terra for the daily work, Luna for volume.
Tier selection FAQ
Not sure which tier? What should the default be?
Make Terra your default: better than GPT-5.5 at half of Sol's price. Escalate to Sol when Terra gets stuck, and drop to Luna for high-frequency light tasks. This Terra-first strategy is the cost optimum for most teams.
Do the tiers share the same context window and output limit?
Yes. OpenAI's docs confirm all three share the 1.05M-token API context window, 128K max output, the same February 16, 2026 knowledge cutoff and the same reasoning-effort options (none through max). Note that clients like Codex apply a lower session cap, as with GPT-5.5's 400K convention.
What is Sol Ultra? Is it billed separately?
Sol Ultra is Sol's high-compute mode (91.9% on Terminal-Bench 2.1), not a separate price tier. It burns more reasoning tokens, so a task costs more at the same rates. When you see 91.9% quoted, note that it is the Ultra-mode score.
How are the tiers billed on QCode?
All three tiers share one QCode key and one quota, billed by usage at the live rates on the /models page. Switching tiers needs no account changes — just use a different model id.
All three tiers are live on QCode
Sign up and call gpt-5.6-sol / terra / luna right away — same key as the Claude and Gemini lineups.