GPT-5.6 Sol vs Terra vs Luna: OpenAI’s New Tier Explained

On June 26, 2026, OpenAI previewed GPT-5.6 as three models rather than one. Sol is the flagship, Terra is the everyday workhorse, and Luna is the fast/cheap tier. This is the same three-tier pattern Anthropic and Google use, but with new reasoning controls and different price/performance tradeoffs.

Last verified: July 1, 2026

The three tiers at a glance

Tier	Positioning	Price (per 1M tokens)	Best for
Sol	Flagship, max reasoning	$5 input / $30 output	Hard problems, cybersecurity, high-stakes code
Terra	Balanced default	$2.50 input / $15 output	Everyday chat, agent loops, mainstream coding
Luna	Fast + affordable	$1 input / $6 output	High-volume classification, summarization, drafts

Sol — the flagship

Sol is OpenAI’s most capable model, period. It’s positioned specifically for:

• Cybersecurity work — OpenAI calls it their “most capable model yet” on ExploitBench, and it was competitive with Anthropic’s Mythos Preview using ~1/3 the output tokens
• Complex reasoning — the new “max reasoning effort” and “ultra mode” controls let you dial up how many tokens Sol burns thinking before responding
• Safety-sensitive workloads — Sol ships with OpenAI’s “most robust safety stack to date,” which is why it’s rolling out gradually under US-government coordination

When Sol is worth $5/$30 pricing:

• The task is hard enough that Terra fails or hallucinates
• You need a defensible reasoning trail for compliance/audit
• Cybersecurity, code review, or research where errors are expensive

Terra — the default

Terra is where most GPT-5.6 usage will land. It’s cheaper than GPT-5.5 while beating it on benchmarks, which is OpenAI’s typical mid-tier refresh pattern.

When Terra is the right pick:

• Chatbot backends, customer support, general-purpose agents
• Coding tasks that don’t need Sol’s reasoning ceiling
• Anything you were running on GPT-5.5 — Terra is the drop-in upgrade

Terra vs Claude Sonnet 5:

• Terra: $2.50 input / $15 output
• Sonnet 5: $2 input / $10 output (intro), $3/$15 after Sep 1
• On paper, Sonnet 5 is slightly cheaper at intro pricing and matches Terra on most benchmarks. The pick often comes down to which ecosystem you’re already in.

Luna — the volume tier

Luna is for tasks where you’re calling the model thousands of times per day. At $1 input / $6 output, it’s designed to compete with:

• Claude Haiku 4.5
• Gemini 3.5 Flash
• Mistral Small
• DeepSeek V4 Flash

When Luna is the right pick:

• Bulk classification (sentiment, category, intent)
• Summarization pipelines
• First-draft generation before a Terra/Sol pass
• Anything where the marginal cost per call needs to be under a penny

New reasoning controls

The GPT-5.6 family introduces two new API parameters:

• reasoning_effort — set to low, medium, high, or max. Higher values burn more tokens on internal deliberation before responding. Available on Sol and Terra.
• ultra_mode — a boolean that pushes Sol into extended reasoning for the hardest problems. Expect noticeably higher latency and cost.

These give developers explicit control over the accuracy/cost/latency triangle instead of relying on model choice alone. Expect this pattern to spread — Anthropic and Google will likely ship equivalents in Q3.

Availability status (July 1, 2026)

Surface	Sol	Terra	Luna
API (trusted partners)	✅	✅	✅
API (broad)	🟡 rolling out	🟡 rolling out	🟡 rolling out
ChatGPT Plus/Pro	🟡 coming weeks	🟡 coming weeks	🟡 coming weeks
Codex	🟡 coming weeks	🟡 coming weeks	🟡 coming weeks

Sol’s limited preview is deliberate: OpenAI is coordinating with the US government on evaluation frameworks for high-capability models, similar to the export-control review that briefly restricted Anthropic’s Fable 5 and Mythos 5 in June.

Picking a tier

Default recipe for a typical AI app:

1. Route by task complexity. Simple classification → Luna. Standard chat/coding → Terra. Hard reasoning or high-stakes output → Sol.
2. Use reasoning_effort as a fallback. If Terra fails on a task, retry with reasoning_effort=high before escalating to Sol.
3. Cache aggressively. GPT-5.6 changed some caching behavior — check the pricing docs for the new discount structure.

When to prefer non-OpenAI alternatives:

• Cost-sensitive agent coding → Claude Sonnet 5 (cheaper than Terra at intro pricing)
• 2M-token context needs → Gemini 3.5 Pro
• Fully open-weights → DeepSeek V4 or Llama 4

The bottom line

GPT-5.6’s three-tier launch is OpenAI catching up structurally to Anthropic’s Opus/Sonnet/Haiku and Google’s Pro/Flash/Flash-Lite splits. The interesting part isn’t the flagship — it’s Terra, which is where most real usage will land, and where Sonnet 5 is already competing hard on price. Choose based on your existing stack, then let usage patterns and cost data drive the tier mix.

---

Last verified: July 1, 2026. Sources: OpenAI’s June 26, 2026 GPT-5.6 Sol preview announcement, finout.io pricing analysis, MarkTechPost coverage of the launch.