GPT-5.6 Sol vs Terra vs Luna: Builders’ Tier Guide

OpenAI’s GPT-5.6 is not one model with three stickers — it is a family of three durable capability tiers: Sol (flagship), Terra (balanced), and Luna (cost-optimized). General availability landed around July 9, 2026. All three share the same generation features and a huge context window; what changes is intelligence headroom vs price, so the real product question is which tier should own each step of your workflow.

This guide is for API builders. Specs and list prices below follow OpenAI’s model pages for Sol, Terra, and Luna. On this platform they are listed as openai/gpt-5.6-sol, openai/gpt-5.6-terra, and openai/gpt-5.6-luna.

What “5.6 + Sol / Terra / Luna” means

Piece	Meaning
5.6	Generation (shared context limits, tool surface, family behavior)
Sol	Frontier tier — hardest coding, research, long-horizon professional work
Terra	Mid tier — everyday production agents; balances quality and cost
Luna	Efficient tier — high-volume, cost-sensitive, routine transformations

OpenAI maps them roughly to earlier GPT-5 naming: Sol ≈ former unsuffixed / flagship slot, Terra ≈ “mini-class” balance, Luna ≈ “nano-class” volume tier — but all three keep the same 1.05M context, which is the important break from older “cheap = tiny context” assumptions.

Shared specs (all three)

Spec	Value
Context window	1,050,000 tokens
Max output	128,000 tokens
Modalities (API model pages)	Text + image in → text out
Long-context surcharge	Prompts with >272K input bill at 2× input and 1.5× output for the full request
Cached input reads	90% off uncached input (e.g. Sol $0.50 / MTok)
Cache writes	1.25× uncached input

So Luna is not “a small-context toy.” It is a cheaper intelligence tier that still accepts the same long prompts — with the same long-context multiplier if you blow past 272K.

The differences that actually matter

1) Price (official list, per 1M tokens)

Tier	API id (OpenAI)	Gateway id (this platform)	Input	Output	Cached input
Sol	`gpt-5.6-sol` (alias `gpt-5.6` → Sol)	`openai/gpt-5.6-sol`	$5.00	$30.00	$0.50
Terra	`gpt-5.6-terra`	`openai/gpt-5.6-terra`	$2.50	$15.00	$0.25
Luna	`gpt-5.6-luna`	`openai/gpt-5.6-luna`	$1.00	$6.00	$0.10

Relative to Sol, Terra is about half the token price; Luna is about one-fifth. Always verify live retail on Pricing — gateway quotes can differ from list.

Trap: the bare alias gpt-5.6 routes to Sol, not Terra. If you “upgrade to 5.6” without picking a tier, you may silently pay flagship rates.

2) Capability / job fit

Use Sol when failure is expensive

Multi-file refactors, deep debugging, research agents, security-sensitive analysis, long plan→execute→verify loops
You need the highest published headroom in the 5.6 family and can afford $5/$30 economics

Use Terra as the default production workhorse

Everyday agents, analysis, coding that is hard but not “bet the company”
OpenAI positions Terra as the balance tier; many teams should start here, not on Sol

Use Luna when volume and $/request dominate

Classification, extraction into schemas, normalization, first-pass summaries, high-QPS transforms with a grader or human review
OpenAI’s own framing: optimized for cost-sensitive, high-volume workloads

Independent reporting after launch also notes Luna can look surprisingly close to Sol on many short coding/text tasks, but can fall off more on hard long-context recall and some agent/computer-use suites — so do not pick Luna for “needle in a 200K+ haystack” without your own eval.

3) Same prompts, different routers

Because the three tiers share the generation’s tool/input surface, a practical architecture is:

Luna for cheap pre-steps (classify, extract, draft)
Terra for the main agent loop
Sol only for escalations (failed checks, high-severity tickets, final synthesis)

That is usually cheaper than “everything on Sol,” and safer than “everything on Luna.”

Quick decision table

Situation	Pick
Interactive IDE agent on a large repo, ambiguous bugs	Sol (or Terra first, escalate to Sol)
Production chatbot / ops agent with tools	Terra
Nightly ETL: 100k extractions / classifications	Luna
>272K tokens of context in one call	Any tier — but budget the 2× / 1.5× long-context rates
“I just set model=`gpt-5.6`”	You are on Sol — confirm that was intentional

How to try all three here

# Flagship
curl https://openfluxhub.com/v1/chat/completions \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-5.6-sol","messages":[{"role":"user","content":"Design a migration plan for this legacy service."}]}'

# Default production
curl https://openfluxhub.com/v1/chat/completions \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-5.6-terra","messages":[{"role":"user","content":"Implement the next slice of the plan and list risks."}]}'

# High volume
curl https://openfluxhub.com/v1/chat/completions \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-5.6-luna","messages":[{"role":"user","content":"Extract JSON fields: company, amount, due_date from this invoice text."}]}'

Tips that save real money across all three:

Keep a stable system + tool prefix so cache reads stay at the 90% discount.
Measure $/successful task, not $/1M tokens in isolation — Sol can win if it finishes in fewer turns.
Watch the 272K threshold; splitting context or retrieving selectively often beats paying the long-context multiplier.
Put Luna behind a grader (schema validation, unit tests, rubric) before you trust it unsupervised.

Bottom line

GPT-5.6’s Sol / Terra / Luna split is a routing problem, not a branding puzzle:

Sol = maximum headroom, highest price (openai/gpt-5.6-sol)
Terra = everyday default for most production agents (openai/gpt-5.6-terra)
Luna = volume and cost, same huge context, less headroom (openai/gpt-5.6-luna)

All three are live on this platform. Compare them on Pricing, run one real workflow through each tier, and lock defaults from measured quality and spend — not from the celestial name alone.