# GPT-5.6 Luna

GPT-5.6 Luna is an AI model released by OpenAI on Jun 26 2026. It has a 1.05M context window. Tracked results include 41% on BullshitBench v2, 59.54% on ARC-AGI-2 and 35.9% on CursorBench 4.0.

## Facts

| Field | Value |
| --- | --- |
| Model | GPT-5.6 Luna |
| Developer | OpenAI |
| Release date | Friday, Jun 26 2026 |
| Licensing | Closed |
| Context window | 1.05M |

## API pricing

All rates in USD per 1,000,000 tokens, pay-as-you-go.

| Tier | Input | Cached input | Cache write | Output |
| --- | --- | --- | --- | --- |
| Up to 272K input tokens | $0.20 | $0.02 | $0.25 | $1.20 |
| Over 272K input tokens | $0.40 | $0.04 | $0.50 | $1.80 |

Verified October 19, 2026 against the first-party source: https://developers.openai.com/api/docs/models/gpt-5.6-luna

## Tracked benchmark scores

| Benchmark | Score | Source | What it measures |
| --- | --- | --- | --- |
| BullshitBench v2 | 41% | [BullshitBench](https://github.com/petergpt/bullshit-benchmark) | Given a confidently-worded but nonsensical prompt, does the AI spot that it makes no sense and push back — instead of playing along and inventing an answer? The score is how often it clearly called out the nonsense. Higher is better. |
| Gray Swan IPI (k = 1) | 8.3% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. Gray Swan's indirect prompt injection benchmark measures how often such an attack succeeds when the attacker gets a single try. Lower is better. |
| Gray Swan IPI (k = 10) | 38.6% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 10 tries and counts an attack as successful if any of them works. Lower is better. |
| Gray Swan IPI (k = 15) | 43.9% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 15 tries and counts an attack as successful if any of them works. Lower is better. |
| CursorBench 4.0 | 35.9% | [CursorBench](https://cursor.com/cursorbench), retrieved 2026-09-28 | Cursor's own test of coding agents on ambiguous, multi-file tasks taken from real Cursor sessions — editing, refactoring, investigating a codebase, understanding what the user meant, managing jobs and following a design. Cursor runs each model at several reasoning efforts; each release here carries the score of its best listed effort. Scores aren't comparable with earlier CursorBench versions. Higher is better. |
| Supabase Evals (with skills) | 63.8% | [Supabase Evals](https://supabase.com/evals) | Supabase's own open benchmark: a coding agent is dropped into a real Supabase project and asked to do real work — set up a schema, fix a broken security policy, debug an Edge Function — and every run is checked against a live Supabase stack. This is the headline number, where the agent has Supabase's own skills loaded, as most people building on Supabase would. The score is the share of scenarios it got right. Higher is better. |
| Supabase Evals (no skills) | 42% | [Supabase Evals](https://supabase.com/evals) | The same Supabase scenarios, but with none of Supabase's skills loaded — so it measures what the model already knows about building on Supabase, rather than how well it follows Supabase's supplied instructions. Higher is better. |
| Frontier-Bench v0.1 | 14.3% | Lab | A hard, ever-evolving set of real computer tasks — coding, system administration, data work, and more — that an AI agent has to complete on its own. Run by the Harbor / Laude Institute team as the successor to Terminal-Bench (v0.1 is the first release of the task set). The score is the share of tasks solved. Higher is better. |
| Terminal-Bench 4.0 | 17.27% | [Terminal-Bench](https://github.com/harbor-framework/terminal-bench), retrieved 2026-08-30 | Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Version 4.0 recalibrated how much time, CPU and memory each task gets, removed eight tasks and fixed nineteen, so fewer runs fail for reasons that have nothing to do with the model. Scores are not comparable with earlier versions. Higher is better. |
| Terminal-Bench 2.1 | 82.5% | Lab | Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Higher is better. |
| BrowseComp | 83.3% | [BenchLM](https://benchlm.ai), retrieved 2026-08-31 | Can the AI browse the web and track down hard-to-find answers? Higher is better. |
| ARC-AGI-2 | 59.54% | [BenchLM](https://benchlm.ai), retrieved 2026-08-31 | Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better. |
| threejseval | 1270 | [threejseval](https://threejseval.com/ranking) | Every model gets the same prompt — "the Eiffel Tower", "a glass fishbowl", "a robot arm picking toys into a box" — and builds a 3D scene in Three.js. Real people then see two scenes side by side, names hidden, and vote for the one they prefer. The votes become a chess-style Elo rating on threejseval.com, averaged across all the prompts. It measures whether the scene looks and moves right to a human eye, not whether the code passes a test. Higher is better. |

## About GPT-5.6 Luna

GPT-5.6 Luna was introduced in the June 2026 family preview as OpenAI's fast, affordable tier for frequent requests. It sat below Terra and Sol, with an emphasis on making useful reasoning available for high-volume workloads. Like its siblings, it began in a restricted preview before the later general rollout.

The API documentation associated Luna with the role previously filled by nano models, while retaining a 1,050,000-token context and adjustable reasoning effort. That gave applications a smaller tier without reducing the family to short-input tasks. Luna's initial positioning, later price changes and eventual successors were separate milestones; this entry records the original GPT-5.6 model rather than transferring later Luna specifications to it.

## Questions and answers

### When was GPT-5.6 Luna released?

GPT-5.6 Luna was released by OpenAI on Friday, Jun 26 2026.

### Who made GPT-5.6 Luna?

GPT-5.6 Luna was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research.

### How much does GPT-5.6 Luna cost?

GPT-5.6 Luna costs $0.20 per million input tokens and $1.20 per million output tokens through the OpenAI API. Cached input is $0.02 per million tokens. Those are the rates for the “Up to 272K input tokens” tier; 1 other pricing tier is published for this model. Rates are pay-as-you-go API prices verified against OpenAI's published pricing on October 19, 2026.

### What benchmark scores did GPT-5.6 Luna get?

GPT-5.6 Luna reports 13 tracked benchmark scores — BullshitBench v2: 41%; Gray Swan IPI (k = 1): 8.3%; Gray Swan IPI (k = 10): 38.6%; Gray Swan IPI (k = 15): 43.9%; CursorBench 4.0: 35.9%; Supabase Evals (with skills): 63.8%; Supabase Evals (no skills): 42%; Frontier-Bench v0.1: 14.3%; Terminal-Bench 4.0: 17.27%; Terminal-Bench 2.1: 82.5%; BrowseComp: 83.3%; ARC-AGI-2: 59.54%; threejseval: 1270. Tracked scores may come from lab reports or independent benchmarks; source details accompany the benchmark data.

### What is the context window of GPT-5.6 Luna?

GPT-5.6 Luna has a context window of 1.05M. Providers may specify separate input and output token limits.

### Is GPT-5.6 Luna open source?

No. GPT-5.6 Luna is a closed model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms.

### What came before and after GPT-5.6 Luna?

OpenAI's previous tracked release was GPT-5.6 Terra on Jun 26 2026. It was followed by GPT-5.6-Cyber on Aug 10 2026.


---

Canonical page: https://aireleasetracker.com/model/openai/gpt-5.6-luna
Site index: https://aireleasetracker.com/llms.txt
Source: AI Release Tracker (https://aireleasetracker.com). Most benchmark scores come from lab launch material; gathered results identify the leaderboard that published them.