Home/Blog/GPT-5.6 Sol vs Terra vs Luna:3ティアの違い、価格、選び方
解説10 分で読める

GPT-5.6 Sol vs Terra vs Luna:3ティアの違い、価格、選び方

OpenAI は 2026 年 7 月 9 日に GPT-5.6 を Sol・Terra・Luna の3ティアでリリースしました。7月30日の値下げ後は価格差が25倍に。ベンチマーク、価格、長コンテキストの限界、用途別の選び方を比較します。

執筆 ZNIX AI Editorial · AI モデルの資料と利用ガイド
公開日

GPT-5.6 Sol vs Terra vs Luna: Differences, Pricing, and Which Tier to Use

OpenAI shipped GPT-5.6 to general availability on July 9, 2026, across ChatGPT, Codex, and the API. It arrives as three models — Sol, Terra, and Luna — and after the July 30 price cuts, the gap between the cheapest and flagship tier is 25× on input and output tokens. The benchmark gap is often much narrower than that. This guide explains the differences, the pricing math, and how to route your work. Figures were checked on November 20, 2026.

What Is GPT-5.6?

The number (5.6) identifies the generation. The names identify what OpenAI calls “durable capability tiers that can advance on their own cadence”: Sol is the flagship, Terra is the balanced everyday model, Luna is the fast, affordable tier. The idea is that Terra can get smarter without becoming Sol, and Luna can get faster without becoming Terra.

In the API, the explicit model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. The bare gpt-5.6 alias routes to Sol.

The release itself was unusual: GPT-5.6 spent 12 days behind a government review gate before going public, under the June 2026 AI cybersecurity order. The Commerce Department's Center for AI Standards and Innovation cleared it after OpenAI sent technical experts to Washington. The review was legally voluntary; practically, it functioned like preclearance.

Sol vs Terra vs Luna at a Glance

DimensionSolTerraLuna
PositioningDeepest reasoning, hardest agentic tasksBalanced everyday defaultSpeed and volume, lowest cost
Input / output per 1M tokens (after July 30)$5 / $30$2 / $12$0.20 / $1.20
Agents' Last Exam53.650.450.3
Terminal-Bench 2.188.8%87.4%84.7%
MRCR long-context recall91.5%89.6%41.3%

Read the last row carefully: Sol and Terra are nearly tied on long-context recall, while Luna falls off a cliff (41.3%). That single number defines where the cheap tier breaks.

Pricing After the July 30 Cuts

Three weeks after launch, OpenAI cut Terra by 20% (from $2.50/$15 to $2/$12) and Luna by 80% (from $1/$6 to $0.20/$1.20). Sol stayed at $5/$30. The cuts went live in the OpenAI API the same day, with AWS Bedrock pricing rolling out alongside.

  • Prompt caching: cache writes cost 1.25× the uncached input rate; cached reads get a 90% discount; cache life is a minimum of 30 minutes with explicit breakpoints. After the cut, Luna cached reads fall to $0.02 per 1M tokens.
  • Batch / Flex processing (already half-rate) prices Luna at $0.10 input / $0.60 output.
  • Fast mode replaced Priority Processing for Sol in the API: up to 2.5× the speed at 2× the standard price, with no change in model intelligence. Existing priority-tagged calls route automatically.

Why Prices Dropped: Sol Rewrote Its Own Inference

The engineering story separates this from a routine pricing move. After GA, OpenAI set GPT-5.6 Sol to work inside Codex on its own production GPU kernels. Sol rewrote those kernels in Triton and Gluon, cutting end-to-end serving costs by 20%, and redesigned its own speculative-decoding draft model through hundreds of autonomous experiments, raising token-generation efficiency by more than 15%. OpenAI validated the rewritten code with FpSan, its open-source floating-point sanitizer.

This is the first publicly documented case of a production frontier model autonomously rewriting its own serving stack and having that rewrite translated into customer pricing.

Benchmarks: Where the Tiers Actually Differ

Long-horizon agentic work. On Agents' Last Exam (55 professional fields), all three tiers beat every non-OpenAI model listed: Sol 53.6, Terra 50.4, Luna 50.3, versus Claude Fable 5 at 40.5. Note the internal spread: Sol to Luna is 3.3 points, while Luna costs 1/25th of Sol.

Terminal and engineering. Sol sets the state-of-the-art on Terminal-Bench 2.1 at 88.8% (single agent); Sol Ultra with four parallel agents reaches 91.9% at roughly 3× the cost. Terra (87.4%) and Luna (84.7%) trail closely. On the Artificial Analysis Coding Agent Index: Sol 80, Terra 77.4, Luna 74.6 — with Fable 5 at 77.2, between Terra and Luna.

The SWE-Bench Pro exception. OpenAI's own tables show Sol at 64.6% versus Claude Fable 5's 80% on repo-level coding. OpenAI published a separate estimate that ~30% of SWE-Bench Pro tasks are broken. Maybe that is right — but when a lab leads a benchmark it cites it, and when it trails it questions the methodology. Until a replacement emerges, Fable 5 remains the strongest public model for pure repo-level coding.

Computer use and browsing. Sol is state-of-the-art on BrowseComp (92.2%) and OSWorld 2.0 (62.6%). Terra surpasses GPT-5.5's peak at lower cost; Luna nearly matches GPT-5.5 at less than half the estimated cost. (Anthropic reports Fable 5 at 85.0% on OSWorld-Verified — a different task set, not directly comparable.)

Long-context recall. MRCR: Sol 91.5%, Terra 89.6%, Luna 41.3%. If your workload is document analysis, large-codebase reasoning, or multi-document synthesis, Luna is the wrong tool.

Max Reasoning and Ultra Mode

GPT-5.6 adds two capability levers on top of the tier system:

  • Max reasoning effort — a new level above “high.” Reasoning effort is configurable per request from “none” to “max” (reasoning.mode: "pro" on any tier), and reasoning now persists across turns instead of restarting each message.
  • Ultra mode — spawns parallel subagents (four by default; the API equivalent is the multi-agent beta in the Responses API). Gains: BrowseComp +1.8, SEC-Bench Pro +3.1, Terminal-Bench +3.1 — at roughly 3× single-agent cost.

Both are quality-first tools for the hardest tasks. For routine work, they waste compute.

How to Get GPT-5.6

  • ChatGPT: Terra and Luna are available in the app; subscription usage now consumes fewer credits (prices unchanged). Auto-review in ChatGPT and Codex CLI was upgraded from GPT-5.4 to Luna, making it roughly 10× less expensive.
  • API: model IDs gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna; the bare alias routes to Sol.
  • Codex and AWS Bedrock: available, with Bedrock pricing aligned to the July 30 cuts.
  • Migration note: GPT-5.4 retired on July 23, 2026; GPT-5.5 stays available. If you are on 5.4, move to Terra or Luna — not Sol — unless you specifically need the flagship tier.

Which Tier Should You Use?

  • Default to Terra. It matches or comes within 2–3 points of Sol on most benchmarks at less than half the price, and beats GPT-5.5's peak on OSWorld and BrowseComp. For most teams this is the answer.
  • Escalate to Sol for the hardest agentic tasks: long-horizon multi-step workflows, complex terminal operations, cybersecurity research, frontier computer use. Max reasoning and Ultra mode are fully supported only on Sol.
  • Use Luna for volume, not depth: classification, summarization, high-volume pipelines. Respect the MRCR ceiling (41.3%) — no long-context recall work.
  • Use Claude Fable 5 for pure repo-level coding while its SWE-Bench Pro lead stands. See our Fable 5 explainer.

What OpenAI Did Not Report

The GPT-5.6 benchmark suite is deliberately agentic: Terminal-Bench, Agents' Last Exam, BrowseComp, OSWorld. Missing: SWE-bench Verified, GPQA Diamond, AIME, MMLU, ARC-AGI-2, FrontierMath — the traditional academic benchmarks. OpenAI also switched to reporting performance curves across reasoning-effort levels instead of single scores.

The charitable read: agentic benchmarks resemble real work better than single-shot academic tests. The cynical read: OpenAI foregrounded the benchmarks where it leads. Both can be true — the full picture requires reading both labs' tables.

Where GPT-6 Astra Fits

On September 3, 2026, OpenAI shipped GPT-6 Astra, the successor to GPT-5.6 Sol. Two honest footnotes from that review: the independent Requesty index scored Astra at 61 — the same as Sol — and Astra's computer use is nearly 2× faster than Sol with a 97.7% agentic safe-completion rate (vs 85.1% for Sol). The tier system introduced with 5.6 is the naming foundation Astra builds on.

For creators, these are reasoning engines that sit upstream of rendering models: script planning, storyboard-to-shot breakdowns, and prompt iteration that feed tools like our AI video generator and AI image generator. They do not replace specialized video models.

Sources and Verification

Figures above were checked on November 20, 2026 against:

Vendor benchmarks are claims, not guarantees. This article is a documentation summary, not a hands-on benchmark.

よくある質問

What are GPT-5.6 Sol, Terra, and Luna?
They are the three capability tiers of GPT-5.6, which OpenAI shipped to general availability on July 9, 2026. Sol is the flagship for the deepest reasoning and hardest agentic tasks, Terra is the balanced everyday model, and Luna is the fast, affordable tier. API model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna; the bare gpt-5.6 alias routes to Sol.
What is the difference between Sol, Terra, and Luna?
Capability, speed, and cost. On Agents' Last Exam the scores are 53.6 / 50.4 / 50.3 — a 3.3-point spread — while after the July 30, 2026 price cuts Luna costs 1/25th of Sol. The sharpest difference is long-context recall (MRCR): Sol 91.5%, Terra 89.6%, Luna 41.3%. Luna is for volume, not for long-context or deep-reasoning work.
How much does GPT-5.6 cost?
After the July 30, 2026 cuts, per million tokens: Sol $5 input / $30 output (unchanged), Terra $2 / $12 (was $2.50 / $15), Luna $0.20 / $1.20 (was $1 / $6). Cache writes cost 1.25x the input rate, cached reads get a 90% discount, and Batch/Flex processing halves rates again. Sol's Fast mode delivers up to 2.5x speed at 2x price.
How do I get GPT-5.6?
GPT-5.6 has been generally available since July 9, 2026 in ChatGPT, Codex, the OpenAI API, and AWS Bedrock, after a 12-day government review under the June 2026 AI cybersecurity order. GPT-5.4 retired on July 23, 2026; GPT-5.5 remains available. Auto-review in ChatGPT and Codex CLI now runs on Luna.
Which GPT-5.6 tier should I use?
Default to Terra: within 2–3 points of Sol on most benchmarks at less than half the price. Escalate to Sol for the hardest agentic tasks, max reasoning, and Ultra mode. Use Luna for high-volume classification and summarization, but never for long-context recall (41.3% MRCR). For pure repo-level coding, Claude Fable 5 still leads SWE-Bench Pro (80% vs Sol's 64.6%).

さらに読む

執筆者について
ZNIX AI Editorial — AI モデルの資料と利用ガイド

ZNIX 編集チームは一次資料へのリンク、日付付きの提供状況確認、明示した例を用いて AI モデルを解説します。資料の要約と実測テストを区別しています。

AI動画を作る準備はできましたか?

登録で50無料クレジット — クレジットカード不要。

無料で始める →