GPT-5.6 Luna is the fast, cost-efficient model in OpenAI’s GPT-5.6 series, tuned for high-volume, latency-sensitive workloads while retaining genuinely capable reasoning. It accepts text, image, and file inputs, returns text, and supports native tool calling, structured outputs, and configurable reasoning effort. You can call it through the GPT-5.6 Luna API model page on OrcaRouter, which documents the pricing and telemetry, or read our GPT-5.6 family pricing guide for how it sits against Terra and Sol.
Most discussion of the GPT-5.6 series centers on the flagship, but most production traffic does not need a flagship. Luna exists for the majority: the chat, classification, extraction, and routing workloads where a cheap, fast model that is still genuinely capable beats an expensive one that is barely better. This article explains what Luna is, what it costs, how it performs, and where it belongs in a production stack.
What makes Luna a workhorse
A workhorse model is one you route most of your traffic to because it is cheap and fast enough to run at volume, and capable enough that you do not pay for it in rework. Luna is positioned exactly there. The model page describes it as tuned for high-volume, latency-sensitive workloads, and the intended uses are concrete: chat, classification, extraction, routing, and lightweight agentic workflows. These are the tasks that make up the bulk of real LLM traffic, and they are precisely the tasks where paying flagship prices is waste.
The price that defines the tier
Luna is priced at $0.20 per million input tokens and $1.20 per million output tokens at the standard tier, with a higher tier at $0.40 in and $1.80 out for longer-context requests. Cache reads cost about $0.02 per million. To put the price in context, the pricing guide notes that Luna undercuts a comparable fast model by roughly five times on input and four times on output, and sits near the floor set by cheap open-weight models. The cost example on the model page is striking: ten million tokens a month at a 70 percent input share comes to about five dollars on Luna. The same workload on the mid-tier runs closer to fifty, and on the flagship closer to one hundred and twenty-five.
The latency that fits interactive products
Luna’s production latency is the other half of its identity. The model page reports a median time to first token of about 1.45 seconds at the p50, and the pricing guide cites a similar figure of about 1.65 seconds. At roughly one and a half seconds to the first token, Luna is fast enough for interactive use — a chat widget, a live classification, an assistant that answers in real time — without the seven-second waits that frontier flagships impose. For products where response time is the product, this is the number that makes Luna usable where a heavier model is not.
The reliability that makes volume safe
The model page reports an error rate of 0.05 percent over the past seven days, during which the platform served roughly 14.9 billion tokens through the model. That is a very low error rate for a very large volume, and it is the property that makes high-volume traffic practical: when a pipeline sends millions of requests, a low error rate means fewer retries, fewer fallbacks, and simpler operations. Reliability is part of the cost story, and on this measure Luna is a quiet standout.
What it can actually do
A low price does not mean low capability. Luna scores 71.4 on the AA coding index, ranked fifteenth of 132 models, and 52.3 on the AA intelligence index, ranked twenty-sixth of 134. It posts 91.1 on GPQA Diamond, 78.3 on long-context recall, and 80.9 on terminal-bench 2.1. Those are mid-field numbers with real competence underneath: it is not the strongest model in the series, and it does not need to be. For chat, extraction, and routing, the capability is more than sufficient, and for lightweight agentic workflows the terminal-bench score says the tool use holds up. The one-million-token context window means it can also handle large inputs without chunking.
Where it sits in the family
Within the GPT-5.6 series, Luna is the bottom rung by price and the one built for volume. Terra sits above it at $2 in and $12 out, the balanced middle for tougher reasoning and coding. Sol sits at the top at $5 in and $30 out, the flagship for deep multi-step reasoning and long-horizon agentic work. The pricing guide’s recommendation is the practical answer: use Luna for the easy majority, Terra when you need more, and Sol only where it earns its cost. The three models share roughly a one-million-token context and up to 128K of output, so the choice between them is about capability and price, not architecture.
When to use it
Luna is the right default for most requests: high-volume chat, classification, extraction, routing decisions, and simple agent steps. Its price makes it possible to run at scale, its latency makes it usable interactively, and its capability handles the routine bulk without rework. If a request is not genuinely hard and does not have a very expensive wrong answer, Luna is likely the correct model.
When not to use it
The cases for a heavier model are the hard tail: complex multi-step reasoning, large-scale software engineering, long-horizon agentic work where a single mistake compounds. For those, Terra or Sol earns its higher price, because the cost of a wrong answer exceeds the token savings. The model’s own positioning is honest about this — Luna is for the volume workhorse role, not for the frontier.
The takeaway
GPT-5.6 Luna is the fast, cost-efficient workhorse of the GPT-5.6 series: $0.20 in and $1.20 out, a median first token around 1.5 seconds, a 0.05 percent error rate, and a one-million-token context, all aimed at high-volume chat, classification, extraction, and routing workloads. It is not the strongest model in the family, and it is not meant to be — it is the model you route the majority of traffic to so the flagship is reserved for the requests that need it. Understanding where Luna belongs is the single biggest cost lever in the GPT-5.6 family.
Sourcing note: Prices, context window, benchmark, and telemetry figures for GPT-5.6 Luna are from the OrcaRouter model page and the Luna/Terra/Sol pricing guide (checked August 2026); benchmarks are sourced from artificialanalysis.ai. Latency and error rate are OrcaRouter production telemetry over the preceding seven days.