GPT-5.6 Sol vs Terra vs Luna: Pricing, Strengths, and How to Start

OpenAI made the GPT-5.6 family generally available on 9 July 2026. The number marks the generation; Sol, Terra, and Luna are capability tiers that can evolve on different tracks. Choosing by name alone is a weak strategy. Practitioners need a clearer map of difficulty, cost, latency, and where the model is actually offered.

📑Table of Contents

1. What GPT-5.6 Is: Sol, Terra, and Luna
2. What Improved: Capability, Efficiency, Agents
3. Strengths by Use Case
4. Comparison Table: Pricing, Fit, Access
5. How to Start: ChatGPT, Codex, API, Copilot
6. Author Perspective
7. FAQ
8. Summary and Next Actions

This article combines OpenAI’s GPT-5.6 GA post and Sol preview post, the GitHub Copilot changelog, and Simon Willison’s practitioner notes. The goal is an adoption-oriented comparison—not a vendor recap. Benchmarks below are vendor-reported reference points as of those posts, not permanent leaderboards.

---

What GPT-5.6 Is: Sol, Terra, and Luna

GPT-5.6 is a family, not a single model SKU.

• Sol: flagship for hard reasoning, long-running agent work, and difficult coding
• Terra: everyday balanced tier; OpenAI positions it as competitive with GPT-5.5-class work at a lower cost band
• Luna: fast and affordable for bulk work, light tasks, and latency-sensitive flows

Limited preview began on 26 June 2026; general availability followed on 9 July 2026. Surfaces include ChatGPT, Codex, and the OpenAI API, with phased rollout. GitHub Copilot also listed Sol, Terra, and Luna on 9 July 2026. In Japan, SUPERNOVA’s Stella AI announced Terra/Luna availability on 10 July 2026 (PR TIMES), a reminder that third-party SaaS can appear quickly after the official GA.

---

What Improved: Capability, Efficiency, Agents

OpenAI’s narrative emphasizes useful work per token and the ability to escalate with max and ultra when needed.

Selected GA claims (vendor-reported):

• Agents’ Last Exam: Sol at 53.6, with comparisons that also stress lower estimated cost at moderate reasoning
• Coding: Sol max reasoning = 80 on Artificial Analysis Coding Agent Index (claimed SOTA); strong Terminal-Bench 2.1 / DeepSWE framing
• BrowseComp: Sol 92.2% (claimed SOTA)
• OSWorld 2.0: Sol 62.6%, with large claimed reductions in output tokens for computer-use style work
• Programmatic Tool Calling on the Responses API: lightweight program execution for tool orchestration, intermediate filtering, fewer round trips
• max / ultra: max for longer deliberation and verification; ultra for multi-agent coordination to reduce time-to-result (multi-agent beta called out for ultra-like setups)

Preview charts also showed Terminal-Bench 2.1 examples such as Sol Ultra 91.9%, Sol 88.8%, Terra 82.5%, and Luna 84.3%. Treat fine-grained gaps carefully: task mix, reasoning effort, output tokens, and harnesses all move rankings.

On safety, OpenAI describes in-model refusals, real-time cyber/bio misuse classifiers, account review, and risk-based access. Cyber capability gains are presented with a defensive safety stack, and OpenAI states the Preparedness Framework Cyber Critical threshold has not been crossed. Higher capability usually means more operational impact from refusals and delayed access.

---

Strengths by Use Case

Pick by quality, latency, cost, and safety boundary—not by novelty.

Priority	Start here	Escalate when
Hard refactors, long agents, high cost of failure	Sol	Only hard slices need max / ultra
Docs, summaries, analysis, everyday coding	Terra	Quality gaps persist on specific tasks
Bulk jobs, light Q&A, speed-first flows	Luna	Quality drops on important work

Practical mapping:

1. Cross-repo design changes → Sol first; max for deeper verification; ultra when parallel agent attack reduces wall-clock time.
2. Meeting notes, draft specs, routine PR text → Terra is usually the sensible default.
3. Log shaping, classification, first-pass document triage → Luna often wins on unit cost and speed.
4. Dual-use cyber-adjacent work → assume refusals and reviews; do not design a fully automated path with no human fallback.

The common failure mode is either under-tiering hard work (endless rework on Luna) or over-tiering easy work (Sol + ultra for every ticket).

---

Comparison Table: Pricing, Fit, Access

API list prices per 1M tokens, as presented across preview/GA materials:

Dimension	Sol	Terra	Luna
Role	Flagship	Everyday balance	Fast / low cost
API price (in / out)	$5 / $30	$2.50 / $15	$1 / $6
Best fit	Hard reasoning, long agents, difficult coding	Standard work, summaries, analysis, daily coding	Bulk processing, light tasks, speed
Escalation tools	max / ultra, multi-agent patterns	Keep cost down while holding quality	Stay in the lowest band
Access examples	ChatGPT (plan-dependent), API, Codex, Copilot (SKU-dependent)	Same family of surfaces	Same family of surfaces

Sources: OpenAI GPT-5.6 GA, previewing GPT-5.6 Sol, GitHub Copilot changelog (9 Jul 2026) (as of July 2026).

Simon Willison notes a shared knowledge cutoff of 2026-02-16, 1M-token context, and 128k max output across the three models. List-price comparisons to Claude SKUs are useful orientation, but reasoning-token behavior means “cheaper per 1M” is not always “cheaper per finished task.”

Prompt caching (preview description): explicit cache breakpoints, minimum ~30-minute cache lifetime, cache writes at 1.25× uncached input, cache reads at a 90% discount. Workflows that resend long fixed prompts will diverge from sticker prices.

Copilot SKU notes from the changelog:

• Sol: Copilot Pro+, Max, Business, Enterprise
• Terra / Luna: Pro and above

Copilot billing rides usage-based billing and provider list pricing—do not assume OpenAI API invoices map 1:1 to GitHub usage.

---

How to Start: ChatGPT, Codex, API, Copilot

OpenAI API

1. Confirm account and billing
2. Smoke-test Sol / Terra / Luna model IDs on a tiny prompt
3. If you need tool orchestration or multi-agent patterns, review Responses API Programmatic Tool Calling and multi-agent beta
4. Design cache breakpoints for long stable prompts
5. Probe refusal-prone domains before production automation

---

ChatGPT / Codex

Availability of max / ultra and specific tiers depends on product and plan. Check admin and help docs first. Codex tends to fit long agentic coding loops; ChatGPT tends to fit interactive knowledge work.

---

GitHub Copilot

Use this path when the IDE is the control plane. Match SKUs before enabling Sol for the whole team, and decide whether finance tracks OpenAI direct spend, GitHub usage, or both.

---

Third-party SaaS (example)

Services such as Stella AI can expose Terra/Luna quickly. Useful for early evaluation if the org already standardizes on that product—but logging, retention, and data handling follow the SaaS contract, not only OpenAI’s defaults.

---

Adoption checklist

• ✓ Map work to Sol / Terra / Luna
• ✓ Run three representative tasks and note quality, latency, and cost
• ✓ Estimate API cost with cache assumptions
• ✓ Choose one access path to start
• ✓ Write an “escalate only for hard slices” rule for max / ultra
• ✓ Add refusal/delay fallbacks for dual-use work
• ✓ Re-check openai.com and GitHub docs before contract or production changes

---

Author Perspective

Chasing the newest flagship for every task is usually the expensive default. A more durable pattern is to park standard work on Terra or Luna and reserve Sol plus higher effort modes for slices where failure is costly. Monthly cost often moves more with cache hit rate, refusal-driven rework, and split billing across API and Copilot than with a one-point benchmark gap.

---

FAQ

Q: What is the difference between Sol, Terra, and Luna?

GPT-5.6 is the generation; Sol/Terra/Luna are capability tiers—flagship, everyday balance, and fast/low-cost.

Q: What does API pricing look like?

As published, Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M tokens (input/output). Cache and reasoning volume change effective cost.

Q: When should I use max or ultra?

Only when the default tier fails quality or completion goals. max favors deeper deliberation; ultra favors multi-agent time-to-result.

Q: Can I use this in ChatGPT immediately?

GA still allows phased availability by product, plan, region, and org contract. Verify in product UI/admin, and fall back to API or Copilot if needed.

Q: Which GitHub Copilot plans include these models?

As of the 9 July 2026 changelog, Sol is on Pro+/Max/Business/Enterprise; Terra and Luna are on Pro and above. Confirm current SKUs in GitHub docs.

Q: Any special caution for cyber-related tasks?

Yes. Expect refusals, delays, and reviews. Keep a human fallback and avoid making automated cyber-adjacent flows a single point of failure.

---

Summary and Next Actions

1. Map tasks to Sol / Terra / Luna
2. Estimate cost with cache assumptions
3. Pick ChatGPT, Codex, API, or Copilot as the first path
4. Escalate only hard slices with max / ultra
5. Operate dual-use domains with refusal/delay expected

Re-check OpenAI’s GPT-5.6 announcement and GitHub Copilot’s changelog/docs before implementation or procurement changes.

Related articles:

• OpenAI GPT-5.6 Late June Release — 1.5M Context and Meaningful Improvements to Rival Claude
• OpenAI Patch the Planet: Accelerating OSS Security Patches with GPT-5.5-Cyber and Codex Security
• OpenAI Announces GPT-5.6: Sol, Terra, Luna Models with US Government Limited Preview

Related new article:

• GPT-5.6 Sol Token Usage and API Cost Estimation Guide (2026) – This published update adds current operational context for GPT-5.6 Sol vs Terra vs Luna: Pricing, Strengths, and How to Start.