Luna replaces expensive per-token inference for high-volume, latency-sensitive workloads like classification and routing, not hard reasoning tasks. It's the only model in the family on ChatGPT's free tier, so most casual users meet it first, while teams needing multi-agent orchestration should upgrade to a pricier sibling.
Luna is OpenAI's efficiency-tier model, released July 9, 2026, with the lowest per-token price and lowest latency of the family, at the cost of a reasoning ceiling capped below Terra and Sol. It targets high-volume inference and consumer ChatGPT Free tier access rather than frontier reasoning work.
Where it sits
- $2.25/M$ per 1M tokensBlended price (3:1)Lower is better#36 / 64peer median $1.70/Mvendor price, checked by HokAI
- 90 tok/stokens/sOutput speedHigher is better#19 / 39peer median 90 tok/scited: Artificial Analysis
- 60.2%% solvedSWE-bench VerifiedHigher is better#25 / 28peer median 78.3%per source, see benchmark scores
- 55.3%% correctGPQA DiamondHigher is better#43 / 44peer median 88.3%per source, see benchmark scores
Priced around the middle of the 64 GA models with a published price (rank 36), in the bottom third on SWE-bench Verified (rank 25 of 28), and one of 22 that document a zero-data-retention option.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT-5.6
Context window: 200,000 tokens · Max output: 64,000
Input modalities: text, image · Output: text, tool-calls, code
About GPT-5.6 Luna
GPT-5.6 Luna is the efficiency tier in OpenAI's GPT-5.6 lineup, which reached general availability on July 9, 2026. It delivers the strongest capability-to-cost ratio of the three, targeting high-volume inference workloads where per-token cost dominates total economics. Luna sacrifices peak reasoning depth and advanced agentic features for raw speed and affordability.
Luna shares the GPT-5.6 architecture family: token efficiency improvements over the prior generation, prompt caching keyed to named breakpoints, persisted reasoning across turns, and vision plus text input with text and tool-calls or code output. It supports reasoning effort levels none, low, medium, and high; xhigh, max, and pro mode stay reserved for Terra and Sol. There is no ultra multi-agent mode and no programmatic tool calling.
The model supports a 200K context window with 64K max output tokens. Native modalities: text and image input; text, tool-calls, and code output. Vision capabilities are present but not benchmarked against dedicated multimodal flagships, and audio or video input/output isn't supported at all. Luna is available on the OpenAI API, ChatGPT on all tiers including Free with rate limits, Codex, Microsoft Copilot with explicit selection, and additional model gateways, with training data current through June 2026.
Head to the pricing FAQ for Luna's exact per-token numbers. Enterprise buyers get Zero Data Retention eligibility, US and EU data residency, plus the standard compliance certifications listed under governance. Latency runs around 800ms p50 and 3000ms p99, the fastest of any GPT-5.6 tier.
Pricing
Input $1/M, Output $6/M, Cached input $0.125/M (1.25x base, 90% discount). 20% of Sol price, 40% of Terra price. No batch discount at launch. ChatGPT Free tier includes rate-limited Luna access. Enterprise ZDR eligible. Data residency: US, EU.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.030 | $0.0060 | $0.036 |
| Support reply | $0.0020 | $0.0018 | $0.0038 |
| One coding agent run | $0.200 | $0.120 | $0.320 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Maximum Cost Efficiency: The cheapest tier in the family, priced for the best token-per-dollar economics on volume workloads.
- Fastest Latency in the Family: Runs faster than both Sol and Terra, making it the pick when p50 latency matters more than peak reasoning depth.
- Full Vision Plus Tool Use: Not stripped down: text and image input, function calling, code execution, and structured output are all present.
- Auto-Caching Plus Persisted Reasoning: Automatic prompt caching carries context forward, and reasoning.context reuses reasoning items across turns.
- ChatGPT Free Tier Access: The only model in the family available on the free ChatGPT tier, which gives it the broadest consumer reach.
Pros
- Unbeatable cost efficiency for high-volume inference, with auto-caching keeping per-request cost near zero at scale.
- The fastest latency in the family, which meets strict SLA requirements for real-time classification, routing, and extraction.
- A full capability surface (vision, tools, code), unlike typical lite models that strip features to hit a price point.
- ChatGPT Free tier inclusion means massive distribution and an upgrade funnel toward the pricier siblings.
Cons
- Reasoning ceiling caps out at high effort, so it can't match Terra or Sol on complex multi-step or novel problems.
- No ultra multi-agent mode, no programmatic tool calling, no pro mode, and no explicit caching control.
- As a distilled model it can show overconfidence on edge cases and subtler reasoning gaps than its teacher models.
- Auto-caching only, with a real write-cost penalty, so volatile prompt prefixes burn budget silently.
Benchmarks
- MATH: 71.5% vendor-reported · 26 Jun 2026 — Competition maths problems, % solved.
- MMLU: 88.9% vendor-reported · 26 Jun 2026 — General-knowledge exam across 57 subjects, % correct.
- MMLU-Pro: 74.8% vendor-reported · 26 Jun 2026 — A harder version of the 57-subject knowledge exam, % correct.
- AIME 2025: 72.1% vendor-reported · 26 Jun 2026 — Competition-level maths problems from the 2025 exam, % solved.
- ARC-AGI 2: 18.7% vendor-reported · 26 Jun 2026 — Abstract visual puzzles built to resist memorisation, % solved.
- HumanEval: 88.4% vendor-reported · 26 Jun 2026 — Small programs that must pass hidden tests, % passing.
- LiveBench: 52.4% vendor-reported · 26 Jun 2026 — A rolling set of fresh questions that cannot have been in training data, % correct.
- LMArena Elo: 1340 independent · 12 Jul 2026 — Rating from blind human votes on which answer is better.
- GPQA Diamond: 55.3% vendor-reported · 26 Jun 2026 — PhD-level science questions that are hard to search for, % correct.
- LMArena rank: #12 independent · 12 Jul 2026 — Position on the blind human-preference leaderboard; #1 is best.
- Aider Polyglot: 62.1% vendor-reported · 26 Jun 2026 — Code edits across several programming languages, % correct.
- SWE-bench Verified: 60.2% vendor-reported · 26 Jun 2026 — Real GitHub issues fixed end to end, % solved.
- Humanity's Last Exam: 14.2% vendor-reported · 26 Jun 2026 — Expert-written questions across many fields, % correct.
- AA Intelligence Index: 45 cited: Artificial Analysis · 10 Jul 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- AA blended price: $3.50/M cited: Artificial Analysis · 10 Jul 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 90 tok/s cited: Artificial Analysis · 10 Jul 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does GPT-5.6 Luna cost per 1M tokens?
Luna's list price is $1/M for input and $6/M for output; cached input runs $0.125/M, a 90% discount off the uncached rate. That's 20% of Sol's per-token price and 40% of Terra's, with no batch discount at launch.
How does GPT-5.6 Luna compare on benchmarks vs GPT-5.6 Sol and Terra?
Luna scores about 60% on SWE-bench Verified and 55% on GPQA Diamond, well behind Terra and further behind Sol on both. It trades that reasoning gap for the lowest price and fastest latency in the family, with no ultra multi-agent mode or programmatic tool calling.
Can you self-host GPT-5.6 Luna, or is it API-only?
Luna is proprietary: OpenAI doesn't publish weights or offer a license to self-host it. It's reachable through the OpenAI API, ChatGPT's free and paid tiers, Codex, Microsoft Copilot as an explicit selection, and select gateway partners.
What does OpenAI do with the data you send to GPT-5.6 Luna?
By default OpenAI doesn't train on API inputs or outputs, and Luna qualifies for Zero Data Retention on enterprise accounts. Compliance-wise it carries SOC2 Type II, ISO 27001, and GDPR coverage plus HIPAA eligibility, with a choice of US or EU data residency.
Who is GPT-5.6 Luna best for and who should avoid it?
Luna fits high-volume production inference like classification, extraction, and routing, plus consumer apps built on ChatGPT's free tier. Avoid it for complex multi-step reasoning, novel problem solving, or any workflow needing multi-agent orchestration; route those to Terra or Sol instead.
HokAI guides covering GPT-5.6 Luna
- Gemini vs ChatGPT as Your Android Assistant: Gemini and ChatGPT can both be set as your Android default assistant, but Android reserves the wake word and screen-reading hooks for only one of them in 2026.