What is GPT 5.6 Luna?
GPT 5.6 Luna is the fast, high-throughput model in OpenAI's GPT 5.6 family, previewed on June 26, 2026 and generally available since July 9, 2026. It runs on Unifically as openai/gpt-5.6-luna. It takes text and image input and returns text, with a 1,050,000-token context window, a 128,000-token max output, and knowledge up to February 16, 2026. Luna is tuned for high-volume and latency-sensitive workloads, and it still holds its own on real work: 50.3 on Agents' Last Exam, ten points above Claude Fable 5, and 74.6 on the Coding Agent Index, above Claude Opus 4.8. Terra sits above it for everyday work and Sol at the top for the hardest problems.
Key features of GPT 5.6 Luna
Beats Fable 5 on Agents' Last Exam
Luna scores 50.3 on Agents' Last Exam, which tests long-running professional workflows across 55 fields. That is ahead of GPT-5.5 at 46.9, Claude Opus 4.8 at 45.2, and Claude Fable 5 at 40.5, from the fastest model in the family.
Built for throughput
Luna is the fastest GPT 5.6 model, made for the calls you fire thousands of times a day. It still scores 74.6 on the Coding Agent Index, above Claude Opus 4.8 at 72.5, and 82.5% on Terminal-Bench 2.1 at max effort.
Reasoning effort you can tune
Luna runs from no reasoning up to max effort. Keep it low for instant replies; raise it when a batch job deserves more thinking. On GeneBench v1 the range moves it from 1.6% to 14.4% pass@1.
Text and image input
Luna reads screenshots and scanned pages alongside short text prompts, so high-volume pipelines can mix the two in one request.
Holds up on CursorBench 3.2
On Cursor's coding-agent benchmark, Luna scores 61.1% at max effort, level with Sonnet 5 at 61.5% and ahead of Composer 2.5 at 56.1%. Strong for the fastest model in the family.
Predictable prompt caching
Explicit cache breakpoints and a 30-minute minimum cache life reuse a stable prompt across a batch instead of resending it on every call. For template-heavy pipelines, that is spend control you set once per run.
Best for
High-volume pipelines
Classification, extraction, summarization, and routing at the family's fastest turnaround.
Latency-sensitive apps
The fast option for chat frontends and agent steps that need instant replies.
Bounded agent steps
Tool-calling subtasks and quick checks inside larger agent loops.
Vision at scale
Screenshots and scanned pages read inline for high-volume document jobs.
Cached repeated prompts
Cache breakpoints reuse a shared template across a whole batch.
Use cases
Route the bulk of your traffic through Luna: ticket tagging, entity extraction, summaries, and short assistant replies where fast turnaround decides the architecture. Use it as the worker model inside agent systems, handling bounded subtasks while Terra or Sol does the planning. It reads screenshots and scanned pages inline, so it fits high-volume document jobs that mix images with short text. It is also a sensible coding helper for draft patches and quick explanations, since it beats the previous Opus generation on terminal work. For template-heavy batch jobs, cache breakpoints reuse the shared prompt across the whole run.
Limitations
Luna trades reasoning depth for speed. On GeneBench v1 it tops out at 14.4% pass@1 versus 28.4% for Terra, and on ExploitBench it reaches 33.2% versus Terra's 52.9%. When a task needs deep multi-step reasoning, move up to Terra or Sol.
GPT 5.6 runs behind layered safeguards. Real-time cyber and biology classifiers can pause or withhold higher-risk generations, and some legitimate dual-use security work may be blocked or slowed.
Arena has not published an Elo for Luna yet. A ranked position is pending while the public rollout is this new.
When to use GPT 5.6 Luna
Use Luna when volume and latency are the constraint: the calls you make thousands of times a day, the agent steps that must return fast, and the batch jobs where turnaround decides feasibility. If output quality on hard reasoning starts to bind, step up to Terra for GPT-5.5-class results; save Sol for the work that needs the strongest model.
API examples
Call GPT 5.6 Luna from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/openai/gpt-5.6-luna.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "openai/gpt-5.6-luna",
"messages": [
{ "role": "user", "content": "Classify this support ticket: billing, bug, or feature request?" }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
openai/gpt-5.6-luna, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
High-volume, latency-sensitive work: classification, extraction, routing, summaries, and short replies where fast turnaround matters most.
For the fastest model in the family, yes. It scores 74.6 on the Artificial Analysis Coding Agent Index, above Claude Opus 4.8 at 72.5, and 82.5% on Terminal-Bench 2.1 at max effort.
1,050,000 tokens, with a max output of 128,000 tokens and a knowledge cutoff of February 16, 2026.
Text and image input, with text output. Good for high-volume jobs that mix short text with screenshots or scanned pages.
Yes, from no reasoning up to max. Low efforts return fast answers; max effort lifts benchmark scores by producing more output tokens.
Luna trades reasoning depth for speed. Terra scores higher on nearly every published eval, and Sol adds max-effort depth plus ultra mode. Luna wins on turnaround and throughput.
GPT 5.6 runs with real-time cyber and biology misuse classifiers. Higher-risk generations can pause for review, and some dual-use security requests may be blocked.