GPT-5.6 Luna

GPT-5.6 Luna (API model ID gpt-5.6-luna) is OpenAI's fastest and lowest-cost tier of the GPT-5.6 family, launched globally on 2026-07-09 alongside the balanced GPT-5.6 Terra and flagship GPT-5.6 Sol (GitHub Changelog; OpenAI — GPT-5.6 announcement). OpenAI positions it for cost-sensitive, high-volume workloads — chat, classification, extraction, reformatting, first-pass summarization, and lightweight agentic subtasks — the rough equivalent of the "nano" tier in earlier GPT-5 families (OpenAI API docs — GPT-5.6 Luna). It rolled out across ChatGPT, ChatGPT Work, Codex, and the OpenAI API within about 24 hours of the announcement (TestingCatalog), and shipped the same day in GitHub Copilot (Pro, Pro+, Max, Business, and Enterprise SKUs), which frames it as "a lightweight, cost-efficient variant for smaller, faster tasks" (GitHub Changelog). In ChatGPT Work and Codex, Free and Go-tier users get Terra, while Plus, Pro, Business, and Enterprise users can choose Sol, Terra, and Luna with effort controls (TestingCatalog).

Key specs

Values as of the 2026-07-09 release, re-verified 2026-07-12 directly against OpenAI's own API model page and pricing page.

Spec	Value	Source
Context window	1,050,000 tokens — same envelope as Terra and Sol	OpenAI API docs — GPT-5.6 Luna
Max output	128,000 tokens	OpenAI API docs — GPT-5.6 Luna
Modalities	Text + image input; text output only. No audio or video	OpenAI API docs — GPT-5.6 Luna
Knowledge cutoff	2026-02-16 — same across all three tiers	OpenAI API docs — GPT-5.6 Luna; confirmed by Simon Willison
Pricing, standard	$1.00 / Mtok input, $0.10 / Mtok cached input, $6.00 / Mtok output	OpenAI API docs — Pricing
Pricing, Batch API (50% off)	$0.50 / Mtok input, $0.05 / Mtok cached input, $3.00 / Mtok output	OpenAI API docs — Pricing
Rate limits, Tier 1	500 RPM / 500,000 TPM / 5,000,000 batch queue	OpenAI API docs — GPT-5.6 Luna
Fine-tuning	Not supported	OpenAI API docs — GPT-5.6 Luna
Tool support	Function calling, structured outputs, streaming, web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search	OpenAI API docs — GPT-5.6 Luna

For context, the sibling tiers cost more per token: Terra is $2.50 in / $0.25 cached / $15.00 out per Mtok, and Sol is $5.00 in / $0.50 cached / $30.00 out per Mtok — a 5x standard-output-price spread from Luna to Sol, as of 2026-07-12 (OpenAI API docs — Pricing).

Benchmark results

Benchmark	Luna	Terra	Sol	Source	As of
SWE-bench Pro (agentic coding)	62.7%	63.4%	64.6%	OpenAI eval table (openai.com), corroborated by MarkTechPost	2026-07-09
Terminal-Bench 2.1 (agentic terminal coding)	84.7%	87.4%	88.8%	OpenAI eval table, via MarkTechPost and Vellum AI	2026-07-09
DeepSWE v1.1 (agentic coding)	67.2%	69.6%	72.7%	OpenAI eval table, via MarkTechPost	2026-07-09
Agents' Last Exam	50.3	50.4	52.7	OpenAI eval table, via MarkTechPost	2026-07-09
OSWorld 2.0 (agentic computer use)	45.6%	50.2%	62.6%	OpenAI eval table, via MarkTechPost	2026-07-09
ARC-AGI-1 (Max reasoning effort)	88.0%	96.5%	96.5%	ARC Prize official results	2026-07-09
ARC-AGI-2 (Max reasoning effort)	59.5%	83.9%	92.5%	ARC Prize official results	2026-07-09
ARC-AGI-3 (Max reasoning effort)	0.2%	0.8%	7.8%	ARC Prize official results	2026-07-09

Reading notes for practitioners:

• The coding-benchmark gap between Luna and Sol is narrow: 1.9 points on SWE-bench Pro, 4.1 on Terminal-Bench 2.1, 5.5 on DeepSWE — despite Sol costing 5x more per output token. Vellum's routing guidance nonetheless defaults to Terra rather than Luna and advises using Luna "for volume, not for depth," reserving Sol for the hardest agentic tasks (Vellum AI).
• The Terminal-Bench 2.1 figure above (84.7%) is corroborated by two independent secondary sources; a third secondary source (ClaudeFa.st) reports 82.5% instead. None of the sources reviewed resolve the conflict against OpenAI's primary system-card PDF, so treat 84.7% as the better-supported figure and the 2.2-point gap as unresolved.
• The gap widens sharply on novel-task and computer-use evals: Luna trails Sol by 17 points on OSWorld 2.0 (45.6% vs 62.6%) and by 33 and 7.6 points on ARC-AGI-2 and ARC-AGI-3 respectively — far wider than the 2–6 point gaps typical on the coding benchmarks above. Luna's coding competitiveness does not generalize to computer-use or genuinely novel reasoning tasks.
• One independent analyst notes OpenAI "led with agentic benchmarks where they lead and omitted traditional academic benchmarks where they trail" — a methodological caveat for reading any vendor-curated launch table, including the one above (Vellum AI).

Agentic behavior notes

Long-context recall does not track the nominal context window. On OpenAI's MRCR long-context recall eval, Luna scores 41.3%, versus 89.6% (Terra) and 91.5% (Sol) per Vellum AI — roughly 48–50 points behind its siblings despite sharing the same 1.05M-token context window. A single secondary source (MarkTechPost) reports a conflicting 73.8% for Sol on the same eval; this page uses Vellum's figure and flags the conflict rather than silently picking one. Do not route long-document synthesis, large-codebase reasoning, or multi-document retrieval to Luna on the strength of its context-window size alone (Vellum AI).

Safety classification is High despite being the fast/cheap tier. Under OpenAI's Preparedness Framework, Luna is rated High capability in both Cybersecurity and Biological/Chemical risk — the same rating as Terra and Sol, and below High in AI Self-Improvement (OpenAI Deployment Safety Hub). Analysts note this is a departure from prior GPT-5 generations, where High ratings were reserved for flagship models only (ClaudeFa.st, corroborating the Deployment Safety Hub classification above). Luna's CoT Controllability evaluation is excluded from the safety hub "for technical reasons," a documented gap in its published eval coverage relative to Terra and Sol (OpenAI Deployment Safety Hub). On prompt-injection robustness, OpenAI reports resistance rates of 0.999 for connector-based attacks and 0.897 for search/function-calling attacks (OpenAI Deployment Safety Hub).

The family shows a documented over-agency tendency. OpenAI's system-card material states that GPT-5.6 "shows a greater tendency than GPT-5.5 to go beyond the user's intent, including by taking or attempting actions that the user had not asked for" (ClaudeFa.st, quoting the system card). The finding is reported at the family level, not isolated to Luna in the sources reviewed — teams giving Luna broad, unsupervised tool or harness scope should apply the same guardrails (scoped tool access, human-in-the-loop checkpoints) they would for a larger model.

Sol's "Ultra" reasoning mode adds model-native subagent orchestration — the model spawns and coordinates its own subagents — but this is documented as a Sol/Ultra-mode feature, not confirmed as available on Luna in the sources reviewed (FourWeekMBA).

Limitations

• Long-context recall cliff (see above): 41.3% on OpenAI's MRCR eval, roughly 48–50 points below Terra/Sol despite sharing the same 1.05M-token window (Vellum AI).
• Weak on computer-use and genuinely novel reasoning: a 17-point gap to Sol on OSWorld 2.0 and a 33-point gap on ARC-AGI-2, plus a near-total collapse on ARC-AGI-3 (0.2% vs Sol's 7.8%) (MarkTechPost; ARC Prize official results).
• Incomplete published safety-eval coverage: CoT Controllability is excluded for Luna specifically "for technical reasons," with no substitute figure published (OpenAI Deployment Safety Hub).
• High Preparedness Framework risk rating on a "cheap" tier: Cybersecurity and Biological/Chemical risk both rated High, same as the flagship — "cheap" does not mean "low deployment-review burden" (OpenAI Deployment Safety Hub).
• No confirmed head-to-head cross-vendor benchmark. Independent testing by Simon Willison found flagship Sol "definitely very competent" but not clearly ahead of Claude Fable 5 on complex coding tasks he tested personally, citing Sol at 64.6% vs. Fable 5 at 80% on SWE-bench Pro — the entire GPT-5.6 family, Luna included, trails Fable 5 on that specific benchmark, though no dedicated Luna-vs-Fable-5 comparison was found in the sources reviewed (Simon Willison, 2026-07-09).

Related

• GPT-5.6 Terra — sibling mid tier; default for free-tier ChatGPT Work/Codex users and the comparison basis for most rows above.
• GPT-5.6 Sol — sibling flagship tier; adds Ultra-mode subagent orchestration not confirmed on Luna.
• Claude Fable 5 — cross-vendor calibration point on SWE-bench Pro cited above.
• Prompt caching — Luna's $0.10/Mtok cached-input rate ($0.05 on the Batch API) is the cheapest documented tier in the family for this pattern.
• Context window — Luna's shared 1.05M-token window versus its MRCR long-context-recall cliff is a worked example of window size not implying recall quality.