GPT-5.6 Luna vs GPT-5.4 Mini
Both models are designed to make capable AI affordable, but they target different points on the quality–latency curve. GPT-5.6 Luna brings a newer generation, a 1.05M context window and stronger published coding and reasoning scores. GPT-5.4 mini is smaller and faster, while Luna is now the cheaper model at current OpenAI API list prices.
Core specifications
| Specification | GPT-5.6 Luna | GPT-5.4 mini |
|---|---|---|
| Provider / release | OpenAI · July 2026 GA | OpenAI · March 17, 2026 |
| API model ID | gpt-5.6-luna | gpt-5.4-mini or gpt-5.4-mini-2026-03-17 |
| Positioning | Cost-efficient GPT-5.6 tier | Strongest GPT-5.4 small model for coding, computer use and subagents |
| Context window | 1,050,000 tokens | 400,000 tokens |
| Maximum output | 128K tokens | 128K tokens |
| Knowledge cutoff | February 16, 2026 | August 31, 2025 |
| Input / output | Text and images / text | Text and images / text |
| Reasoning effort | High tier in the GPT-5.6 family | None, Low, Medium, High and XHigh |
| Tools | Web, file, code, shell, computer use, MCP and tool search | Web, file, code, shell, computer use, MCP, tool search and skills |
| License | Proprietary API | Proprietary API |
This is the cleanest comparison in the set because both models come from OpenAI and the API feature surface is similar. The important distinction is tier: Luna is a newer, more capable model; mini is optimized for throughput and delegation. A well-designed agent can use both rather than forcing one model to handle every subtask.
Benchmark scoreboard
These numbers are from OpenAI's published model guides. Scores use the stated effort levels and benchmark versions. Where the version differs, the table says so instead of presenting a false direct comparison.
| Evaluation | Luna | GPT-5.4 mini | Reading |
|---|---|---|---|
| SWE-Bench Pro, public | 62.7% (GPT-5.6 scorecard) | 54.4% (XHigh) | Luna +8.3 |
| GPQA Diamond | 92.3% | 88.0% | Luna +4.3 |
| Toolathlon | 53.4% | 42.9% | Luna +10.5 |
| MMMU Pro, no tools | 78.4% | 76.6% | Luna +1.8 |
| Terminal-Bench | 84.7% on 2.1 | 60.0% on 2.0 | Version mismatch; Luna result is not a clean head-to-head |
| OSWorld | 45.6% on OSWorld 2.0 | 72.1% on OSWorld-Verified | Different evaluation names and protocols |
| MCP Atlas | Not published | 57.7% | Mini-only published row |
| HLE with tools | Not in Luna scorecard | 41.5% | Mini-only published row |
Radar: capability versus context
● Luna ● GPT-5.4 mini
The first four axes use published percentages. Context is normalized to the 1.05M Luna window: mini is approximately 38% of Luna's context capacity. It is not a quality score.
- Luna's advantage is clearest on coding, GPQA and tool use.
- Mini stays close on MMMU Pro and is specifically optimized for rapid subagent work.
- The context gap is much larger than the price gap.
Why GPT-5.4 mini remains compelling
Speed and delegation
OpenAI positions GPT-5.4 mini as a fast model for coding assistants, subagents, screenshot interpretation and high-volume multimodal applications. It runs more than twice as fast as GPT-5 mini and uses only 30% of GPT-5.4's Codex quota. That makes it a natural worker model behind a larger planner: search a codebase, process supporting documents, classify outputs, or perform narrow fixes in parallel.
Luna's quality headroom
Luna's higher SWE-Bench Pro, GPQA, Toolathlon and long-context results indicate a wider safety margin when a task becomes ambiguous or multi-step. Its 1.05M context also accommodates larger repositories and research bundles without immediately switching to compaction or retrieval.
| Workflow | Better fit | Why |
|---|---|---|
| High-volume classification or extraction | GPT-5.4 mini | Fast response profile; Luna is now cheaper on token rates. |
| Subagents doing bounded supporting work | GPT-5.4 mini | OpenAI explicitly recommends it for delegation. |
| Difficult repository repair | Luna | 8.3-point SWE-Bench Pro lead in the cited scorecard. |
| Large codebase or document context | Luna | 1.05M versus 400K context. |
| Screenshot-driven computer use | Test both | Mini has a strong OSWorld-Verified row, but Luna's row uses OSWorld 2.0. |
| Tool-rich general agent | Luna | Higher Toolathlon score and larger reasoning margin. |
Pricing: Luna is now cheaper than GPT-5.4 mini
OpenAI's current API documentation lists the following standard prices per 1M tokens. OpenAI has reduced Luna's standard input and output prices by 80% versus the rates previously shown in this article: Luna is now cheaper on both input and output, while retaining the larger context window.
| Per 1M tokens | GPT-5.6 Luna | GPT-5.4 mini |
|---|---|---|
| Fresh input | $0.20 | $0.75 |
| Cached input | $0.02 | $0.075 |
| Output | $1.20 | $4.50 |
| Example: 10M input + 1M output | $3.20 | $12.00 |
| Context-specific price tier | Above 272K: long-context rate applies | No separate long-context surcharge listed in the model guide |
At current standard rates, Luna costs about 73% less than GPT-5.4 mini for the same mix of 10M input and 1M output tokens. Its input and output rates are each 26.7% of mini's rates. Batch and Flex prices are lower for both models, but Luna remains the cheaper option; regional-processing uplifts and tool charges may also affect the final bill.
Long context and computer use
The models are not simply “large” and “small” versions of the same experience. Luna's 1.05M window makes it suitable for repository-scale context, long research trails and large reference bundles. GPT-5.4 mini's 400K window is still generous for normal applications and helps keep memory and latency lower, but it will require more aggressive retrieval or compaction on very large tasks.
Conversely, mini's official OSWorld-Verified result of 72.1% shows why smaller models can be strategically better for interactive computer use: the job often rewards fast screenshot interpretation, tool reliability and short feedback loops more than maximum abstract reasoning. The correct production test should include screenshot resolution, click policy, retry budget and end-to-end time—not only the model score.
Which one should you deploy?
Choose GPT-5.6 Luna if…
- Your task has difficult code reasoning, ambiguous requirements or long multi-file context.
- You want the stronger published scores across SWE Pro, GPQA and Toolathlon.
- You want a million-token context without moving to a different model family.
- The lower Luna token rates make capability and long context affordable at high volume.
Choose GPT-5.4 mini if…
- You need fast, latency-sensitive responses for high-volume product traffic.
- You are building a planner/worker system with many narrow subagents.
- You process screenshots or computer-use tasks where speed matters and the context fits under 400K.
- You want to keep a strong GPT-5.4 tool surface with fast response times.
Measure the trade-offs on your own prompts
Run a shadow evaluation with the same tools, context, acceptance tests and latency budget. The 8.3-point SWE-Bench difference is a starting hypothesis, not your product's final conversion rate.
Compare them in CodingFleet →Final verdict
GPT-5.6 Luna is now the stronger value choice. It leads the directly comparable official rows, offers more than twice the context, has a newer knowledge cutoff and is currently cheaper on standard API token rates.
GPT-5.4 mini remains the latency and delegation choice. It is designed for high-volume subagents, runs fast and remains strong on multimodal and computer-use workflows. The OSWorld-Verified result is a reminder that a smaller model can be the right model for a fast interactive loop.
For most systems, start with Luna when cost and capability are the priorities, and use mini where lower latency, bounded subagents or interactive computer use justify the trade-off. Let real success-rate and latency telemetry decide the routing threshold.
Sources and methodology
- OpenAI — GPT-5.6. Luna benchmark scorecard and pricing context.
- OpenAI — Introducing GPT-5.4 mini and nano. Mini benchmark tables, subagent positioning and availability.
- GPT-5.6 Luna API documentation and GPT-5.4 mini API documentation. Context, modalities, tools, effort controls and token rates.
- OSWorld-2.0 and OSWorld-Verified are shown separately because benchmark names and protocols differ. All other provider-reported rows retain their stated effort and version. The radar is a visual summary, not a universal score.