GPT-5.6 Sol vs Terra vs Luna: Which Tier Should You Actually Pick? (July 2026)
The GPT-5.6 Family at a Glance
Announced by OpenAI in July 2026, the GPT-5.6 family splits capability into three explicit tiers instead of the single-model “reasoning effort” slider approach used in earlier versions.
| Tier | Input / MTok | Output / MTok | Position | Typical use |
|---|---|---|---|---|
| Sol Ultra | $12.50 | $75 | Frontier | Terminal agents, multi-agent, hardest reasoning |
| Sol | $5 | $30 | Flagship | Planning, tool use, long-horizon workflows |
| Terra | $2.50 | $15 | Balanced | Everyday coding, RAG, writing |
| Luna | $1 | $6 | Efficiency | High-volume, latency-sensitive, batch |
OpenAI’s own framing: “Sol is the smartest, Terra is competitive with GPT-5.5 at about half the cost, Luna is the fastest and cheapest but least capable.”
Benchmark Reality
| Benchmark | Sol Ultra | Sol | Terra | Luna |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 91.9% | ~85% | ~78% | ~65% |
| SWE-bench Pro | ~70% | ~66% | ~60% | ~48% |
| DeepSWE 1.0 | ~60% | ~57% | ~52% | ~40% |
| MMLU Pro | ~88% | ~86% | ~82% | ~74% |
| HLE | ~53% | ~50% | ~45% | ~35% |
| ARC-AGI-2 | ~65% | ~62% | ~55% | ~44% |
Quality drops per tier: ~3-5 percentage points from Sol Ultra → Sol, ~5-7 from Sol → Terra, ~8-12 from Terra → Luna.
Cost drops per tier: 2.5x from Sol Ultra → Sol, 2x from Sol → Terra, 2.5x from Terra → Luna.
For most workloads, the cost curve is steeper than the quality curve. That’s the whole argument for tiered routing.
The Cost-Per-Task Comparison
A typical 30K-input / 5K-output coding task:
| Model | Blended cost |
|---|---|
| Luna | $0.06 |
| Grok 4.5 | $0.09 |
| Sonnet 5 (intro) | $0.11 |
| Terra | $0.15 |
| Gemini 3.1 Pro | $0.12 |
| Sol | $0.30 |
| Sol Ultra | $0.75 |
| Opus 4.8 | $0.83 |
Luna is the new cheap-tier price floor. Sol Ultra sits at the very top of the market alongside Opus 4.8.
Which Tier for Which Job?
Pick Sol Ultra when:
- Running autonomous terminal agents (Codex CLI, Claude Code equivalents) with 30-100+ tool calls
- Multi-agent orchestration where the top-of-leaderboard delta compounds
- Hardest debugging or refactoring where a single mistake wastes hours of work
- Not for: RAG, standard coding, writing, batch tasks
Pick Sol when:
- Long-horizon planning across many steps
- High-stakes tool use where reliability matters more than cost
- Complex reasoning that’s not quite Sol Ultra territory
- Not for: Batch generation, most RAG, simple code edits
Pick Terra when:
- Daily coding (this is the sweet spot)
- Standard RAG and document synthesis
- Writing where GPT-5.5-level quality is fine
- Cost-sensitive workflows where Luna’s quality drop is too risky
- Not for: Hardest agent work, benchmark-topping tasks
Pick Luna when:
- High-volume, latency-sensitive tasks (autocomplete, in-app AI features)
- Batch code generation where 85% of Sol’s quality is acceptable
- Prototyping and experimentation
- Simple classification/extraction
- Not for: Multi-hop reasoning, complex tool use, mission-critical output
The Multi-Model Routing Pattern
The GPT-5.6 tier split is designed for routing. A representative 2026 stack:
Task → Model
├── Autocomplete → Luna
├── Batch coding → Luna
├── Daily coding → Terra
├── Code review → Terra
├── RAG synthesis → Terra
├── Long-context (1M+) → Claude Sonnet 5 or Gemini 3.5 Pro
├── Terminal agent → Sol Ultra
├── Multi-agent → Sol
└── Hardest reasoning → Sol Ultra or Opus 4.8This routing captures 5-10x cost savings vs “always use Sol Ultra” while preserving quality where it matters.
Luna vs Grok 4.5 vs Sonnet 5
At the cheap end of the market, three models converge:
| Model | Cost (30K/5K task) | Best at |
|---|---|---|
| Luna | $0.06 | Simple tasks, high volume, latency-sensitive |
| Grok 4.5 | $0.09 | Agentic coding (Cursor collab), speed |
| Sonnet 5 (intro) | $0.11 | Reasoning, code review, 1M context |
Rough guide: Luna wins pure cost per token, Grok 4.5 wins agentic speed, Sonnet 5 wins reasoning depth. For output-heavy workflows (long generation), Luna’s $6/MTok output is the killer number.
Migration From GPT-5.5
If you’re on GPT-5.5 today:
- Direct replacement → Terra (OpenAI describes it as GPT-5.5-competitive at half the cost)
- Upgrade path → Sol if you need better tool use / planning
- Downgrade path → Luna if quality is over-provisioned
If you’re on GPT-5.5 Ultra or higher tiers:
- Direct replacement → Sol Ultra for terminal agents, Sol for general use
- Cost cut → Sol saves ~40% vs GPT-5.5 Ultra with similar-tier quality
The Bottom Line
GPT-5.6’s four tiers are the clearest cost/quality menu OpenAI has ever shipped:
- Sol Ultra — top of the market, price it accordingly
- Sol — flagship default for hard work
- Terra — the quiet hero, best cost-adjusted value for daily use
- Luna — new price floor, use for high-volume simple work
Route by task type, not by “always use the smart one.” A $10K/month AI budget spent 60% on Luna + 30% on Terra + 10% on Sol/Sol Ultra will beat $10K spent entirely on Sol on most real workloads.