GPT-5.6 Sol vs Terra vs Luna: Which Tier to Use (2026)

TL;DR

• GPT-5.6 went generally available July 9, 2026 in three tiers: Sol (flagship), Terra (balanced default), and Luna (cheapest, fastest)
• Per 1M tokens: Sol is $5/$30, Terra is $2.50/$15, Luna is $1/$6 — a 5× spread from top to bottom
• The benchmark gap is much smaller than the price gap: Terra lands within ~2 points of Sol on most tests at half the cost, so Terra is the default for most work
• Go to Sol only for the hardest agentic and long-horizon tasks; use Luna for high-volume short-context jobs — but not long-context (its recall collapses)

OpenAI didn't ship one GPT-5.6. It shipped three — Sol, Terra, and Luna — at prices that span 5× from top to bottom. The obvious question for anyone calling the API: do you actually need the flagship, or is the cheap tier good enough? The honest answer is that the benchmark gap is far narrower than the price gap, so most workloads should not default to Sol. Here's the price-versus-performance breakdown, tier by tier, with the one place the cheap tier really does fall apart.

What are GPT-5.6 Sol, Terra, and Luna?

GPT-5.6 Sol, Terra, and Luna are three tiers of the same model family, released together on July 9, 2026, that trade capability for cost. Sol is the flagship, Terra is the balanced everyday tier, and Luna is the cost-and-speed champion.

They share the same generation — Programmatic Tool Calling (the model writes JavaScript that runs in an isolated V8 sandbox with no network access) and the new predictable prompt caching apply across all three. What changes between tiers is raw capability, latency, and price. All three are available now through the OpenAI API, ChatGPT, Codex, and GitHub Copilot.

GPT-5.6 tiers at a glance:

*Source: OpenAI GPT-5.6 launch, July 9, 2026. gpt-5.6 is an alias that points at the flagship.*

The framing OpenAI pushes — "intelligence that scales with your ambition" — really means: pick the smallest tier that clears your task's bar, and only pay up when you have to.

How much does each GPT-5.6 tier cost?

Per 1M tokens, Sol costs $5 input / $30 output, Terra costs $2.50 / $15, and Luna costs $1 / $6. That's a clean 2× step down at each tier and a 5× total spread between Sol and Luna.

Token prices are abstract until you run them against a real workload. Take a moderate production app doing 10M input and 2M output tokens a month:

Monthly cost for a 10M-in / 2M-out workload:

*Illustrative math on published per-token rates; your split of input vs output will shift the totals.*

Same task, 5× difference in the bill. That gap is the whole reason to care which tier you route to — and why defaulting every call to Sol is the most common way teams overspend on this family.

How do GPT-5.6 Sol, Terra, and Luna compare on benchmarks?

On most benchmarks the three tiers cluster far closer than their prices suggest — Terra typically lands within about 2 points of Sol, and Luna within a few more. The exception is long-context recall, where Luna falls off a cliff.

Here are published numbers across the family:

GPT-5.6 benchmark comparison:

*Source: published GPT-5.6 benchmark tables, 2026 (Vellum, OpenAI). Sol also posts 92.2% on BrowseComp; Sol Ultra hits 91.9% on Terminal-Bench 2.1.*

Two things jump out. First, on coding and agentic tasks the tiers are close — Luna holds roughly 85% of Sol's quality at one-fifth the price, which is an excellent trade for high-volume work. Second, MRCR long-context recall is where the cheap tier breaks: Luna drops to 41.3% while Sol and Terra stay near 90%. If your workload feeds long documents or large contexts, Luna is off the table regardless of price.

When should you use each GPT-5.6 tier?

Default to Terra, escalate to Sol only for the hardest agentic work, and drop to Luna for high-volume short-context jobs. Don't pick a tier by prestige — pick it by what the task actually demands.

• Terra — the default. It matches or comes within ~2 points of Sol on most benchmarks at half the price. For most production apps, chat, coding assistance, and general agent steps, Terra is the right starting point. Reach for something else only when a specific task proves Terra isn't enough.
• Sol — the hard cases. Long-horizon multi-step workflows, complex terminal operations, cybersecurity research, computer use, and anything needing max reasoning or ultra mode. Sol is the only tier that fully supports the frontier features, and it leads every benchmark — just not by enough to justify running it everywhere.
• Luna — volume, not depth. High-throughput, latency-sensitive pipelines where cost per query matters more than the last few points of reasoning: classification, tagging, routing, short summaries, bulk generation. Keep it away from long-context work.

The practical pattern most teams land on isn't "pick one." It's routing: Luna or Terra for the bulk of calls, Sol reserved for the steps that genuinely need it. That's where a mixed-tier setup pays for itself.

How do you switch between GPT-5.6 tiers in the API?

You switch tiers by changing the model ID — gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna — while the rest of the request stays identical. Because the tiers share one interface, routing by task is a one-field change.

from openai import OpenAI

client = OpenAI(api_key="sk-...")  # use an environment variable, never hardcode

def answer(prompt, hard=False, bulk=False):
    # route by task: cheap by default, escalate only when needed
    model = "gpt-5.6-sol" if hard else "gpt-5.6-luna" if bulk else "gpt-5.6-terra"
    return client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
    )

The catch with going direct is that this routing lives in your code, tied to one provider's tiers. If you later want to send some traffic to a cheaper non-OpenAI model, or fall back when a tier is rate-limited, you rebuild it. A unified gateway like Velokey puts Sol, Terra, Luna, and other vendors' models behind one OpenAI-compatible endpoint, so tier routing and cross-model fallback become config, not a rewrite — which is the whole point of splitting a workload across tiers to control cost.

For the launch context and how GPT-5.6 got here, see our GPT-5.6 release coverage; for how it stacks up against other frontier models, our GLM-5.2 vs GPT-5.5 vs Claude Opus 4.8 breakdown and the Claude API pricing guide are useful next reads.

Frequently Asked Questions

Which GPT-5.6 tier should I use by default?

Terra. It performs within about 2 points of Sol on most benchmarks at half the price ($2.50/$15 per 1M tokens versus $5/$30). Start production workloads on Terra and only escalate to Sol for tasks that measurably need more — long-horizon agents, complex terminal work, or max-reasoning modes.

What's the price difference between Sol, Terra, and Luna?

Per 1M tokens: Sol is $5 input / $30 output, Terra is $2.50 / $15, and Luna is $1 / $6. That's a 2× step down at each tier and a 5× spread from Sol to Luna. On a 10M-in/2M-out monthly workload, that's roughly $110 (Sol) versus $22 (Luna).

Is GPT-5.6 Luna good enough for production?

For high-volume, short-context tasks, yes — Luna holds around 85% of Sol's quality at one-fifth the price. But its long-context recall drops to 41.3% (versus ~90% for Sol and Terra), so avoid Luna for long documents or large-context prompts. Use it for classification, tagging, routing, and bulk generation.

What are the GPT-5.6 API model IDs?

gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. The bare gpt-5.6 alias points at the flagship (Sol). Switching tiers is a one-field change in the request, since all three share the same API interface.

When is Sol worth 5× the price of Luna?

When the task genuinely needs frontier reasoning: long-horizon multi-step agents, complex terminal operations, cybersecurity research, computer use, or long-context work where Luna's recall fails. For everyday chat, coding help, and routine agent steps, the benchmark gain over Terra or Luna rarely justifies Sol's cost.

Can I mix GPT-5.6 tiers in one application?

Yes, and most cost-efficient setups do. Route the bulk of calls to Luna or Terra and reserve Sol for the hardest steps. Since the tiers share one interface, you can select the model per request. A unified gateway makes this routing (plus fallback to other models) a config change instead of a code change.

Want to route across GPT-5.6 tiers without hardcoding one provider? Get your Velokey API key and call Sol, Terra, and Luna — plus other vendors' models — through one OpenAI-compatible endpoint, with per-request model selection and automatic fallback.

*Last updated: July 10, 2026. Pricing and benchmark figures are from OpenAI's GPT-5.6 launch and published evaluation tables at the time of writing. Numbers are set by OpenAI and may change — verify current rates and benchmarks before committing to production budgets.*