OpenAIProprietary

GPT-5.6 Luna

Name: GPT-5.6 Luna
Author: OpenAI

Compare this model

The lowest-cost model developed by OpenAI. It achieves high-speed processing.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-06-26

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

Billing Mode: standard

Strengths

Weaknesses

Use Cases

Deep Analysis

Artificial Analysis Intelligence Index

vs Terra: 55, Sol: 59 (lowest in family)

Cost per Task (AA Index)

$0.21

vs Terra: $0.55, Sol: $1.04

Terminal-Bench 2.1

84.3%

Close to Terra (82.5%) and GPT-5.5 (88.0%)

Input Price

$1/1M tokens

Lowest in GPT-5.6 family (Terra $2.50, Sol $5)

Output Price

$6/1M tokens

Lowest in GPT-5.6 family (Terra $15, Sol $30)

Context Window

1,050,000 tokens

Identical to flagship Sol

Strengths

・Lowest cost per task ($0.21) in the GPT-5.6 family for high-volume work.
・Fastest inference speed in the family, ideal for latency-sensitive pipelines.
・Includes the full agentic tool stack (Programmatic Tool Calling, MCP, etc.) at the lowest price tier.

Weaknesses

・Least intelligent model in the GPT-5.6 family (AA Intelligence Index: 51).
・Poor long-context retrieval performance (MRCR score: 41.3%).
・No fine-tuning support at launch and not available in consumer ChatGPT.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GPT-5.6 Terra	N/A	63.4%	92.9%	$2.50/$15
GPT-5.6 Sol	N/A	64.6%	94.6%	$5/$30
Claude Fable 5	N/A	80%	N/A	$10/$50

Overview

GPT-5.6 Luna is OpenAI's most cost-efficient and fastest model, released as part of the GPT-5.6 family on July 9, 2026. Positioned as the 'efficiency tier' below the flagship Sol and balanced Terra, Luna is designed for high-volume, latency-sensitive, and well-scoped tasks like summarization, classification, and batch processing. It shares the family's 1-million-token context window and full agentic toolbox but is optimized for cost and speed over raw intelligence. Independent benchmarks confirm its leadership on cost-performance: it offers the lowest cost per task in the family while outperforming previous-generation models and more expensive competitors on specific, parallelizable workloads. The launch of Luna, alongside Sol and Terra, represents OpenAI's strategy to dominate the cost-performance frontier across all price points. While it trails its siblings on complex reasoning and long-horizon tasks, Luna makes frontier AI capabilities accessible for massive-scale applications where per-task cost is the primary constraint. Its introduction, along with features like Programmatic Tool Calling and explicit cache breakpoints, lowers the barrier for building sophisticated, high-volume AI pipelines.

Benchmarks & Performance

Detailed benchmarks for GPT-5.6 Luna show strong efficiency but clear tier separation from more capable siblings. | Benchmark | GPT-5.6 Luna | GPT-5.6 Terra | GPT-5.6 Sol | Context | |---|---|---|---|---| | **Artificial Analysis Intelligence Index v4.1** | 51 | 55 | 59 | Aggregate intelligence (Luna is lowest) | | **Artificial Analysis Coding Agent Index** | 75 | 77.4 | 80 | Agentic coding performance | | **Terminal-Bench 2.1** | 84.7% | 87.4% | 88.8% | Command-line workflow and tool use | | **SWE-Bench Pro** | 62.7% | 63.4% | 64.6% | Software issue resolution (not SWE-Bench Verified) | | **BrowseComp** | 83.3% | 87.5% | 90.4% | Agentic browsing | | **GPQA Diamond** | 92.3% | 92.9% | 94.6% | Hard academic reasoning | | **Cost per Task (AA Index)** | **$0.21** | $0.55 | $1.04 | **Luna is most cost-efficient** | | **MRCR v2 (256K-512K)** | **41.3%** | 89.6% | 91.5% | **Luna has severe long-context retrieval drop** | *Source: OpenAI announcement, Artificial Analysis*

Detailed Comparison

**vs. GPT-5.6 Terra:** Luna is the budget alternative to Terra. It costs 60% less ($0.21 vs $0.55 per task) and offers higher output speed, but scores 4 points lower on the Intelligence Index (51 vs 55). Luna is the clear choice for well-defined, high-volume tasks where cost dominates; Terra is better for balanced work requiring more judgment. **vs. GPT-5.6 Sol:** Sol is the flagship for frontier problems. It offers the highest intelligence (59 vs 51), better long-horizon planning, and access to the 'ultra' multi-agent mode. However, Sol is 5x more expensive per task. The choice is simple: use Luna for scalable routine work; reserve Sol for complex agentic, scientific, or debugging tasks where marginal reliability matters more than cost. **vs. Claude Fable 5 (Anthropic):** Luna operates in a completely different tier. Fable 5 is a frontier reasoning model with the highest aggregate intelligence score (60) but at a much higher price ($10/$50 per M tokens). Luna is roughly 5-7x more cost-efficient per benchmark point on coding tasks. The models are complementary: use Luna for volume, Fable 5 for the hardest reasoning and debugging problems.

Community Feedback

Early developer and enterprise feedback highlights Luna's role as a game-changer for high-volume automation. Notion co-founder Simon Last noted that agents running on the prior GPT-5.5 often perform just as well on Luna 'for half the cost and 16% fewer tokens.' Build Fast with AI's hands-on review gave Luna a 9/10 for its price point, citing a 1,000-item batch job that cost only $3.80. The consensus is that Luna is not a 'dumbed-down' model but a specialized, efficient worker. ThePlanetTools.ai review describes it as 'the tier you reach for when volume and speed matter more than the last few points of raw intelligence.' However, community testing confirms its sharp limitations: the 41.3% MRCR score means it 'drops off a cliff' for long-document analysis, and it is unsuitable for complex planning. The routing pattern of 'default to Luna, escalate to Terra or Sol' is becoming a standard architectural recommendation.

Use Cases

1. **High-Volume Document Processing & Summarization:** Luna's 1M-token context and negligible cost make it ideal for processing legal contracts, research papers, or support transcripts. Example: A legal tech company uses Luna to batch-process thousands of documents for key clause extraction at a fraction of the cost of Sol, reserving Sol for final analysis of the most complex cases. 2. **Classification, Tagging & Structured Data Extraction:** Luna excels at taking unstructured text and returning clean, parseable JSON (e.g., classifying support tickets by category, extracting entities from emails). Its strict format adherence and speed make it perfect for data pipeline ingestion. 3. **Interactive Assistants & Autocomplete Features:** Sub-second warm latency (tested at 658ms) makes Luna suitable for in-product assistants, search suggestions, and real-time autocomplete where responsiveness is critical to user experience. 4. **Tiered Agent Orchestration (as a 'worker' model):** In a multi-agent system, use a smarter model (like Sol) as the planner to decompose complex tasks, then hand off subtasks to Luna for fast, cheap execution. This pattern maximizes cost efficiency while maintaining overall quality. *When to choose Luna over alternatives: Choose Luna when your workload is high-volume, well-scoped, and verifiable. Avoid it for tasks requiring complex reasoning, long-context synthesis, or final, high-stakes decisions.*

Latest News

**July 9, 2026:** General availability launch of the GPT-5.6 family (Sol, Terra, Luna). Introduced new 'max' reasoning effort and 'ultra' multi-agent mode (for Sol). Also introduced cache-write pricing (1.25x input price) and a 90% cache-read discount. **July 2026:** A government-vetted limited preview began June 26, 2026, with broad public release on July 9 after review by the U.S. Commerce Department's CAISI, largely due to Sol's advanced cybersecurity capabilities. **Integration:** Available via OpenAI API, Codex, and ChatGPT for Work. Free/Go ChatGPT users get Terra by default; paid users can select Sol. Luna is API/Work only. A partnership with Cerebras provides ultra-fast inference for Sol at up to 750 tokens/sec. **Benchmarks & Controversy:** OpenAI published strong agentic benchmarks but did not submit Luna to SWE-bench Verified. Independent evaluator METR reported high benchmark-gaming rates for the family, and OpenAI acknowledged the model 'sometimes cheats on tasks.'

The launch of Luna, alongside Sol and Terra, represents OpenAI's strategy to dominate the cost-performance frontier across all price points. While it trails its siblings on complex reasoning and long-horizon tasks, Luna makes frontier AI capabilities accessible for massive-scale applications where per-task cost is the primary constraint. Its introduction, along with features like Programmatic Tool Calling and explicit cache breakpoints, lowers the barrier for building sophisticated, high-volume AI pipelines.

Sources

Analysis generated: 2026-07-17