GPT-5.6 Luna Review: Specs, Pricing & Benchmarks

GPT-5.6 Luna is the fastest and cheapest model in OpenAI's GPT-5.6 family, released on July 9, 2026 alongside GPT-5.6 Sol and Terra. It was not built to win benchmarks. It was built to get through a large volume of ordinary work quickly, at a price low enough that you stop counting tokens.

This review covers what GPT-5.6 Luna is good at, where it clearly is not, what it costs after OpenAI cut its price by 80%, and how it compares with the rest of the GPT-5.6 line. It is also one of the models you can use free inside Chat Smith, with no OpenAI account needed.

GPT-5.6 comes in three tiers: Sol as the flagship, Terra for everyday work, and Luna as the most cost-efficient of the three. OpenAI describes Luna as roughly equivalent to the nano tier in earlier GPT-5 families, aimed at cost-sensitive, high-volume jobs like classification, summarisation, routing, and real-time applications.

The specs are not entry-level, though. Luna has a 1,050,000-token context window, returns up to 128,000 tokens, takes both text and image input, and has a knowledge cutoff of February 16, 2026. It is a reasoning model with effort settings running from none through low, medium (the default), high, xhigh, and max — so the same model can be dialled from near-instant replies to slow, careful work.

At launch OpenAI reported that Luna outperformed Claude Opus 4.8 on coding while taking roughly a third of the time, producing about half as many output tokens, and costing around a quarter as much. That performance-per-dollar framing is the entire point of this tier.

Four things define what Luna is for.

Luna costs $0.20 per million input tokens and $1.20 per million output tokens, with cached input at $0.02. That follows an 80% price cut on July 30, 2026, down from the $1 and $6 it launched at. At this rate, work that would be uneconomical on a flagship model — tagging every support ticket, summarising a whole document queue, routing incoming requests — becomes routine.

Most cheap models give you one speed. Luna's effort setting runs from none to max, so the same integration can answer a simple classification instantly and still spend real thinking time on the occasional hard case. Quality scales with effort rather than forcing a switch to a different model mid-pipeline.

Luna is not stripped down on capability surface. It supports web search, file search, code interpreter, image generation, computer use, MCP, structured outputs, and function calling, and it reads images as well as text. Having the same tool access as the larger tiers at this price is what makes it viable as the default model inside an agent stack.

The numbers hold up better than the price suggests. Luna scored 62.7% on SWE-Bench Pro and 84.7% on Terminal-Bench 2.1, both ahead of GPT-5.5 at a fraction of the cost. On Agents' Last Exam, which tests long professional workflows, it reached 50.3% — within 2.4 points of flagship Sol. For drafting content, summarising long text, and routine analysis, most people would not notice the gap to the top tier.

OpenAI's published figures for the three GPT-5.6 tiers, with the previous generation included for reference.

Read the last row carefully. Luna keeps pace with the larger tiers on most tasks, then collapses on long-context recall. That is the clearest line between this tier and the ones above it.

Luna is the entry point to the GPT-5.6 family on price, sitting at one-tenth of Terra's input rate.

• API: $0.20 input, $0.02 cached input, and $1.20 output per million tokens, available as gpt-5.6-luna.
• ChatGPT Work and Codex: Plus, Pro, Business, and Enterprise users can select Luna and set its effort level.
• Cloud platforms: also served through Microsoft Foundry on Azure and Amazon Bedrock.
• Long prompts: requests above 272K input tokens are billed at 2x input and 1.5x output for the whole request.
• Chat Smith: GPT-5.6 Luna is free to try, with no separate account or API key.

Long context is the big one. Despite the million-token window, Luna scored 41.3% on OpenAI's MRCR 8-needle test at both the 256K–512K and 512K–1M ranges, against 91.5% and 73.8% for Sol. The window is there, but recall across it is not dependable — hand it a 500-page contract and expect it to miss things.

Specialist depth drops off too. On GeneBench Pro it managed 10.8% against Sol's 28.7%, and on FrontierMath Tier 4 it scored 58.5% against 83.0%. Work at the hard end of mathematics, genomics, or research engineering belongs on a larger model.

Computer use is the third gap: 45.6% on OSWorld 2.0 against 62.6% for Sol. If you want a model to drive software end to end rather than answer questions about it, GPT-6 Astra is the model built for that job.

You can use GPT-5.6 Luna free inside Chat Smith — no OpenAI account, no API key, no extra subscription. Pick it from the model list and start typing, on Android, iOS, or in the browser.

Because Luna is quick, it works well as the model you leave open for everyday questions, short rewrites, and summaries, then switch away from when a task needs more depth. In Chat Smith it sits in the same speed class as Gemini 3.5 Flash and Claude Haiku 4.5, so you can send the same prompt to all three and see which style suits you. Context carries over when you switch mid-conversation.

A simple rule: Luna for volume and speed, a flagship model for anything long, technical, or high-stakes.

GPT-5.6 Luna is the model to reach for when the work is ordinary and there is a lot of it. It lands within a couple of points of the flagship on professional and coding benchmarks, runs faster, and costs a fraction as much — and it gives that back on long-context recall, computer use, and specialist depth. For everyday chat, drafting, and summarising, that trade is close to invisible.

You can try GPT-5.6 Luna free in Chat Smith and switch to a heavier model the moment a task calls for it.