"What's the actual difference between Sol, Terra, and Luna? Which one should I use for everyday coding? And are max and ultra the same thing?"
After GPT-5.6 shipped, the first question developers face usually isn't "should I upgrade?" but how to route across the three tiers. OpenAI changed the naming rules this time: the number represents the generation, while Sol / Terra / Luna represent stable capability tiers across generations — you'll see combinations like "GPT-5.7 Sol" and "GPT-6 Terra" going forward. Think in tiers when choosing, not re-learning model names every release.
This article focuses on comparison and selection: how the three tiers differ, which to use in which scenario, how to configure reasoning effort, and how the new caching rules save money. For launch background and the government-review timeline, see our full GPT-5.6 overview.
1. Three Models: Positioning & Pricing
Start by putting all three tiers' prices and positioning side by side. Prices are per 1 million tokens, with input and output billed separately — remember one thing: output is always 5–6× more expensive than input. The main savings lever is producing less output and picking the right tier.
| Model | Model ID | Positioning | Input (/1M) | Output (/1M) |
|---|---|---|---|---|
| Sol | gpt-5.6-sol |
Flagship: strongest to date, step-change gains in coding / biology / cybersecurity; supports max deep reasoning and ultra subagents | $5.00 | $30.00 |
| Terra | gpt-5.6-terra |
Balanced daily driver: performance close to GPT-5.5 at half the cost | $2.50 | $15.00 |
| Luna | gpt-5.6-luna |
Fast & economical: OpenAI's lowest-cost tier; Luna also got significantly stronger after cybersecurity reasoning-tier improvements | $1.00 | $6.00 |
New naming rule: generation × tier
OpenAI's own words: "In this new naming system introduced with GPT-5.6, the number represents the model generation, while Sol, Terra, and Luna represent stable capability tiers that can each evolve on their own schedule." In other words: Terra is the everyday default ($2.50/$15), Sol is for hard problems ($5/$30), Luna is for lightweight wrap-up ($1/$6) — tier names stay stable across generations, so you don't have to re-learn your routing every major release.
2. Scenario Selection (Three-Step Method)
The core principle in one sentence: default to Terra — OpenAI describes it as "performance close to GPT-5.5 at half the price," the sweet spot for price / quality. Sol is for critical tasks Terra can't handle; Luna is for clearly simple work.
| Model | Positioning | Typical Tasks |
|---|---|---|
| Terra | Everyday default | Routine coding, knowledge work, customer support, general agentic tasks — use it when you're not sure |
| Sol | Hard problems | Difficult coding (Terminal-Bench 2.1 SOTA), biology workflows, vulnerability research and exploitation, long-horizon planning |
| Luna | Lightweight | Classification, summarization, batch processing, large-scale cheap inference, latency-sensitive simple tasks |
Three-step method (just follow it):
- Default to Terra — don't overthink it; send your main workload to Terra.
- Clearly simple / high-volume / latency-sensitive? Classification, summarization, formatting, million-row batch jobs → drop to Luna.
- Terra still isn't enough / long-horizon complex agentic work? Only then call in Sol, optionally with
maxeffort orultramode.
Don't "start cheap and escalate tier by tier"
Send your main workload to Terra directly — using Luna and then redoing the work on a higher tier often costs more. Only call Sol when Terra isn't enough — the same mental model as Claude's "Opus day to day, Fable for hard problems." If you're already using our Claude model selection guide, the logic transfers directly.
3. What Changed vs GPT-5.5
If you're migrating from GPT-5.5, these five points are the core changes:
| Dimension | GPT-5.5 | GPT-5.6 |
|---|---|---|
| Naming | Single name: GPT-5.5 | Number (generation) + Sol/Terra/Luna (capability tiers), tiers stable across generations |
| Everyday price | Full GPT-5.5 price ($5/$30) | Terra at roughly half ($2.50/$15), comparable performance |
| Reasoning effort | low / medium / high / xhigh | Sol adds max — the most generous thinking time, for the hardest problems |
| Collaboration mode | Single agent | Introduces ultra mode — subagents work in parallel, accelerating complex work |
| Prompt caching | Implicit caching, unpredictable behavior | Explicit cache breakpoints + 30-minute minimum lifetime; writes at 1.25×, reads at 0.10× |
Official benchmark highlights (Sol tier):
- Terminal-Bench 2.1 (command-line workflows: planning / iteration / tool coordination) — Sol sets a new SOTA
- GeneBench v1 (long-horizon genomics and quantitative biology) — Sol beats GPT-5.5 using fewer tokens
- ExploitBench (cybersecurity vulnerability research and exploitation) — Sol competes with Mythos Preview using roughly 1/3 the output tokens
- ExploitGym (UC Berkeley + OpenAI collaboration) — all three tiers (Sol / Terra / Luna) improve significantly as reasoning effort increases
4. Reasoning Effort & ultra Mode
GPT-5.6 keeps the reasoning-intensity tiers from the GPT-5 family, with Sol adding max — these two things are often confused, so let's separate them first:
| Tier | Best For | Cost Tendency |
|---|---|---|
minimal |
Lowest tier supported by Codex CLI — almost no thinking, answers directly | Extremely cheap |
low |
Classification, simple queries, latency-sensitive non-reasoning work | Cheapest |
medium |
Cost-sensitive routine requests that still need decent quality | Balanced, lean |
high |
Recommended — the sweet spot for most reasoning-sensitive work | Recommended |
xhigh |
Complex coding / long agentic tasks | Quality-leaning |
max NEW · Sol |
Hardest problems; correctness >> cost | Most expensive |
Easy to confuse: max ≠ ultra
- max is a reasoning-intensity tier — lets a single agent think longer, best for deep reasoning that can't be parallelized.
- ultra is a collaboration mode — subagents work in parallel to accelerate complex work, best for multi-path tasks that can be split (multi-angle review, multi-path exploration).
- Official data: Sol Ultra scores 91.9% on Terminal-Bench 2.1, above base Sol's 88.8%.
Refer to the OpenAI API documentation for exact parameter syntax. Codex CLI everyday setup examples:
- Hard problems:
codex -m gpt-5.6-sol -c model_reasoning_effort="max" - Everyday: set
model = "gpt-5.6-terra"+model_reasoning_effort = "high"in~/.codex/config.toml, then just runcodex
5. New Prompt Caching Rules
Caching billing and behavior changed for GPT-5.6 and all future new models. Three things to remember at once:
| Item | Rule |
|---|---|
| Explicit cache breakpoints | Mark cache boundaries explicitly in the prompt, replacing the old implicit behavior — you control exactly which segment goes into cache |
| Cache lifetime | Minimum 30 minutes — unlike the old version that could expire in seconds |
| Write pricing | Cache writes billed at 1.25× the model's uncached input rate |
| Read pricing | Cache reads get a 90% discount off uncached input (billed at 10%) |
Using Terra's $2.50/M input as an example:
- Uncached input (first call): $2.50
- Cache write (first call with breakpoint): $2.50 × 1.25 = $3.125
- Cache read (hit within 30 minutes): $2.50 × 10% = $0.25
Bottom line: as long as a long prompt (system prompt, reference docs, codebase) gets hit ≥ 3 times within 30 minutes, you're already ahead — writes cost 25% more, reads cost only 10%. The higher the hit rate, the more you save. Longer-context scenarios (RAG, long agent sessions) get the biggest leverage.
6. Access Paths: API / Codex / ChatGPT
| Channel | Current Status | How to Switch Models |
|---|---|---|
| API | Rolling out from July 9 | Specify model: "gpt-5.6-terra" etc. in the request body |
| Codex CLI | Opening in sync | /model picker · codex -m gpt-5.6-sol · config.toml |
| ChatGPT | Rolling out in batches, not all at once | Model picker will show Sol / Terra / Luna tiers (depends on your account) |
| Cerebras Fast | July, select customers | Sol up to 750 tokens/second, for latency-sensitive scenarios |
If you don't see the three tiers in ChatGPT yet, that's normal — OpenAI is using a "gradual expansion" strategy. Watch your account's model menu or official announcements.
7. Cost Optimization Cheat Sheet
Three main savings levers for the GPT-5.6 family:
| Lever | How | Where You Save |
|---|---|---|
| Pick the right tier | Default Terra, simple work on Luna, call Sol only when necessary | Avoid over-provisioning |
| Use explicit caching | Add cache breakpoints to long system prompts / reference docs / codebases | Input side |
| Lower reasoning effort | Use low / medium for non-reasoning work; only enable max when correctness >> cost | Output side |
| Compress output | Structured output (JSON schema), explicit "conclusions only / code only", word limits | Output side |
| Batch API | Latency-tolerant tasks via Batch API (long-standing 50% discount policy) | 50% off all tokens |
Three quick savings moves
- Default to Terra, don't escalate to Sol
- Add cache breakpoints to long prompts
- Start reasoning at high, not max
That's the biggest savings leverage. For a more systematic cost-cutting approach, see our LLM pricing and selection guide and Claude Code cost-cutting recap.
FAQ
Which tier should I use day to day — Sol, Terra, or Luna?
Default to Terra. When unsure, use it; simple / high-volume work goes to Luna; only escalate to Sol when Terra isn't enough.
Are max and ultra the same thing?
No. max is single-agent deep reasoning; ultra is subagent parallel collaboration. Different dimensions — can be used together.
What are the prices for each tier?
Sol $5/$30, Terra $2.50/$15, Luna $1/$6 (per 1M tokens, input/output).
When does explicit caching pay off?
The same long prompt hit ≥3 times within 30 minutes is usually already worth it. Terra cache reads cost only $0.25/M.
How do I switch models in Codex?
/model, codex -m gpt-5.6-terra, or persistent config in ~/.codex/config.toml.
Conclusion
How to choose GPT-5.6 Sol, Terra, or Luna? One sentence: Terra by default, Sol for hard problems, Luna for wrap-up — pick by capability tier; the number is just a generation marker.
Start everyday work on Terra (half the price of GPT-5.5, comparable performance); only call Sol for the hardest problems (with max deep reasoning or ultra subagents); hand all clearly simple work to Luna. Don't escalate tier by tier — sending your main workload to Terra directly is the most hassle-free GPT-5.6 selection approach in 2026.
Running GPT-5.6 agents on iOS / macOS builds? Pair with a stable Cloud Mac
If Sol's ultra subtasks touch Xcode or TestFlight, you need a stable macOS execution node. Vuncloud's dedicated Mac mini M4, available on monthly rental, keeps your agent's overnight builds online.
Related Reading
- What Is GPT-5.6? Sol, Terra & Luna Family, Pricing and Government-Reviewed Launch Explained (2026)
- How to Choose the Latest Claude Models? Sonnet 5, Fable 5, Opus 4.8 Compared
- 2026 LLM Pricing, Config, Performance & Who Should Use What
- Codex Weekly Limit Exhausted? 7 Fixes, Quota Mechanics & Alternative APIs (2026)
Pricing and features per OpenAI's official blog post, pricing page, and Codex documentation. Last updated: July 10, 2026.