HomeAI models
OpenAI: GPT-5.6 Luna
OpenAI
•
GPT-5.6 Luna (max)
•
Released 
July 2026

OpenAI
GPT-5.6 Luna (max)

OpenAI's fastest, lowest-cost GPT-5.6 tier for high-volume classification, routing, and lightweight agentic workflows at scale.

Modality:
Text
Image
PDF
model ID
openai/gpt-5.6-luna
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

117.57
tok/s

Intelligence Index *

37.5
/ 100

Context Window *

1050000
tokens

Input price

1.2
Anytoken

Output price

7.2
Anytoken
GPT-5.6 Luna: The Cost-Efficient GPT-5.6 Tier for High-Volume Inference GPT-5.6 Luna is the fastest and most affordable model in OpenAI's three-tier GPT-5.6 family, sitting below Sol (flagship) and Terra (balanced). OpenAI positions it as the successor to the nano tier from earlier GPT-5 generations, built for cost-sensitive, high-volume workloads. Luna accepts text and image input and returns text, offers a large context window, and includes configurable reasoning effort. Its strongest fit is per-request work where unit economics dominate: classification, routing, summarization, and real-time applications where latency and token cost matter more than maximum accuracy. Start sending high-volume traffic to GPT-5.6 Luna through the AnyAPI.ai API.

Performance

Where Luna Delivers: Throughput and Cost per Task

Luna's defining strength is capability per dollar on high-volume work. Independently, it scores 38 on the Artificial Analysis Intelligence Index at max reasoning, well above the median for its price tier, and generates output around 115 tokens per second. In practice this means Luna can absorb per-request features—classification, routing, extraction—that stay economically viable only when the cost per call remains low. Because reasoning effort is configurable from none through max, teams can dial the model toward speed for simple tasks or toward accuracy for harder ones, tuning throughput and cost without switching model tiers.

Benchmarks

Luna on Independent Benchmarks: Fast, Cheap, and Surprisingly Capable

On the Artificial Analysis Coding Agent Index, Luna scores 74.6, trailing Terra (77.4) and Sol (80) but landing above Anthropic's Opus 4.8 at 72.5. On Terminal-Bench 2.1 it reaches roughly 84.7%, within a point of the previous-generation GPT-5.5 at a fraction of the cost. The important caveat is TTFT and long context: at max reasoning, independent measurement shows time-to-first-token of around 128 seconds, and third-party testing reports long-context recall degrading to roughly 41%. Luna is strong on high-throughput classification and short-context coding, but weak where latency or long-document recall is critical.

Output Speed

*
117.57
tok/s

Intelligence Index

*
37.5
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
91
%

HLE *

Adherence to multi-step structured instructions.
40
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

GPT-5.6 Luna accepts text and image input and returns text only; it is not an audio or image-generation model. It pairs a 1,050,000-token context window with up to 128,000 output tokens per request, so long inputs and long outputs are handled independently. The most production-relevant details are the long-context pricing multiplier—prompts above 272K input tokens are billed at higher input and output rates—and configurable reasoning effort (none through max), which directly controls latency and cost. Prompt caching and structured JSON-schema outputs are supported, and the model is OpenAI-SDK compatible via Chat Completions and Responses APIs.
Verified Specifications — 
GPT-5.6 Luna (max)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1050000
 tokens
Maximum output tokens
128000
Reasoning
Yes
Knowledge cutoff
July 2026
Pricing (standard)
1.2
 AnyTokens in
 / 
7.2
 AnyTokens out

Comparison

GPT-5.6 Luna vs GPT-5.6 Terra: When to Move Up a Tier

Luna and Terra are adjacent tiers in the same GPT-5.6 family and share the 1.05M-token context window, reasoning controls, and API surface, so switching between them is a configuration change rather than a re-architecture. The realistic decision is economic and reliability-driven: Luna is the fastest and cheapest tier, while Terra is the balanced production default. On the Artificial Analysis Coding Agent Index they are close (74.6 vs 77.4), but Terra pulls meaningfully ahead on harder agentic and long-context tasks such as OSWorld and multi-turn trajectories, where Luna's recall limitations become a hard constraint.

Dimension
GPT-5.6 Luna (max)
GPT-5.6 Terra (max)
Context window *
1050000
tokens
1050000
tokens
Output speed *
117.57
tok/s
119.18
tok/s
Intelligence Index *
37.5
42.3
Input pricing
1.2
AnyToken
12
AnyToken
Output pricing
7.2
AnyToken
72
AnyToken
Knowledge cutoff *
July 2026
July 2026

Choose GPT-5.6 Luna when request volume dominates cost and each task is short and easily verified—classification, routing, extraction, first-pass generation—and you can tolerate higher time-to-first-token under heavy reasoning. Choose GPT-5.6 Terra when tasks span long documents, require multi-turn agent trajectories, or carry higher error cost, since Terra offers lower latency to first token and stronger long-context behavior. Many teams run both: Luna handles the high-volume path, Terra handles the harder cases the classifier escalates.

Limitations & Trade-offs

Where GPT-5.6 Luna (max) falls short

1
High time-to-first-token under reasoning. Independent measurement at max reasoning shows TTFT around 128 seconds, and even lower effort levels add latency before the first token appears. For interactive chat or streaming UIs where perceived responsiveness matters, this is a genuine constraint; Terra offers far lower time-to-first-token (a few seconds). Use Luna's lower reasoning-effort settings, or a different tier, for latency-critical front-ends.
2
Long-context recall degradation. Despite the 1,050,000-token window, third-party testing reports recall collapsing to roughly 41% on long-context retrieval. The large window makes Luna appear suited to long-document work, but accuracy does not hold across it. For long-document RAG, contract analysis, or multi-turn agents that accumulate context, prefer Terra or Sol rather than routing long inputs to Luna under cost pressure.
3
Trails siblings on hard coding and agentic tasks. Luna scores 74.6 on the Artificial Analysis Coding Agent Index versus Terra's 77.4 and Sol's 80, and drops further on demanding evaluations like OSWorld (about 45.6% vs Sol's 62.6%). It is capable for straightforward, easily verified coding steps in high-volume pipelines, but is not the right tier for complex repository work or long-horizon engineering agents.
4
Long-context pricing multiplier. Prompts exceeding 272K input tokens are billed at 2x input and 1.5x output for the full request, which erodes Luna's core cost advantage precisely when inputs grow large. Combined with weaker long-context recall, this makes very large prompts a poor fit; keep Luna requests within its efficient short-to-moderate context range for the best economics.

Best-Fit Workloads

Where this model earns its place

01

High-volume classification and routing

‍
Luna's low per-request cost and above-median intelligence for its tier make it well suited to classifying support tickets, moderating content, and routing queries at scale. These tasks are short, easily verified, and run on every request, so unit economics dominate—exactly the workload OpenAI and Amazon Bedrock position Luna for. Set reasoning effort low to keep latency and cost minimal.

02

Structured extraction from documents

‍
With JSON-schema structured outputs and image input, Luna can extract fields from invoices, forms, and PDFs into deterministic structures for downstream systems. Keep individual requests within a moderate context to avoid the long-context recall drop and the 272K-token pricing multiplier. For pipelines processing millions of documents, Luna's cost profile makes per-document AI viable where higher tiers would not be.

03

High-throughput coding agents in CI/CD

‍
On Terminal-Bench 2.1 Luna lands near 84.7%, within a point of the prior-generation GPT-5.5 at a fraction of the cost, making it a strong fit for coding agents in CI/CD pipelines where cost-per-run compounds across many executions. Route the high-volume, easily verified steps to Luna and escalate genuinely hard cases to Terra or Sol.

04

Real-time, cost-sensitive generation

‍
For first-pass drafting, autocomplete-style suggestions, and lightweight agentic steps where speed and cost outweigh maximum accuracy, Luna's ~115 tokens/second output and low price tier fit well. Keep reasoning effort low to minimize latency, and note that time-to-first-token rises sharply at higher effort, so reserve heavy reasoning for non-interactive paths.

Pricing in anytokens via AnyAPI
Input
1.2
₳
Output
7.2
₳
Cache write
1.5
₳
Cache read
0.12
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsComparisonLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access GPT-5.6 Luna (max) via AnyAPI.ai

Access GPT-5.6 Luna (max) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate GPT-5.6 Luna (max) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access GPT-5.6 Luna (max) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test GPT-5.6 Luna (max) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use GPT-5.6 Luna (max) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use GPT-5.6 Luna (max) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

GPT-5.6 Luna is OpenAI's fastest, most affordable GPT-5.6 tier, built for cost-sensitive, high-volume workloads. It fits classification, routing, summarization, structured extraction, and lightweight agentic or real-time tasks where cost per request and throughput matter more than maximum accuracy. It is not the right choice for long-document work or complex, long-horizon agents.

GPT-5.6 Luna has a 1,050,000-token context window and supports up to 128,000 output tokens per request. Note that prompts exceeding 272,000 input tokens are billed at a higher multiplier, and independent testing shows long-context recall degrading significantly, so very large prompts are not Luna's strongest use case despite the large window.

Yes. GPT-5.6 Luna is a reasoning model with configurable effort levels—none, low, medium (default), high, xhigh, and max—that control the trade-off between speed and accuracy. It supports function calling, structured JSON-schema outputs, and prompt caching, and through the Responses API it adds web search, file search, and code execution.

Luna is faster and cheaper; Terra is the balanced production default with lower time-to-first-token and stronger long-context and agentic performance. On the Coding Agent Index they are close (74.6 vs 77.4), but Terra pulls ahead on harder tasks. Use Luna for high-volume simple work and escalate to Terra for long-context or complex agents.

Yes. GPT-5.6 Luna accepts both text and image input and returns text output. It can analyze and describe images and process files such as PDFs alongside text. It does not generate images or audio—output is text only.

* Benchmark data source: Artificial Analysis artificialanalysis.ai