GPT-6 Luna review, pricing and verdict

Our verdict: GPT-6 Luna is a strong buy for high-volume extraction, classification and chat, and a weak pick for coding agents. Artificial Analysis's Coding Agent Index slipped two points to 41 against the previous Luna, and its max-effort runs waited 74 seconds for a first token. Choose GPT-6 Sol when code quality decides the job.

GPT-6 Luna, released by OpenAI on 22 September 2026, is cheap by current standards: OpenAI halved its per-token rate against the previous Luna at launch. Artificial Analysis measured $0.07 to run each Intelligence Index task, with a 1,050,000-token context window and six reasoning-effort levels behind that price.

Provider: OpenAI · Family: GPT-6

More about OpenAI on Hokai

Context window: 1,050,000 tokens · Max output: 128,000

Input modalities: text, image · Output: text

About GPT-6 Luna

GPT-6 Luna, unveiled alongside an updated GPT-6 Sol on September 22, 2026, is the smallest and cheapest tier in OpenAI's GPT-6 lineup. Both arrived about three weeks after GPT-6 Astra, OpenAI's September 3, 2026 flagship, and OpenAI describes Sol and Luna as extending Astra's alignment work down to a faster, lower-cost tier. Luna sits below Sol in the lineup: Sol targets interactive and agentic coding, while Luna targets high-volume, latency-sensitive work like summarizing documents, extracting information and answering quick questions, with enough headroom at higher reasoning effort to take on lighter software-engineering and computer-use tasks that used to require Sol.

Is it cheap? Yes. OpenAI's launch post halved the list price against GPT-5.6 Luna (the full rate card is in the Pricing section), and the saving survives contact with real work: Artificial Analysis's cost per task for its Intelligence Index run fell about 60% from the $0.18 the older model needed, even though Luna wrote more output tokens per task (51,000 against 41,000).

OpenAI has not published traditional academic scores (SWE-bench, GPQA, MMLU) for Luna specifically. Its own launch results, reported by Digital Applied on 22 September 2026, cover a set of agentic evaluations at max reasoning effort: 20.7% on AutomationBench, 50.9% on Agents' Last Exam, 42.4% on FrontierCode 1.1, 66.6% on DeepSWE 1.1, 52.7% on OSWorld 2.0 (offline), and a 7.6% factual error rate, down from GPT-5.6 Luna's 8.5%. Independently, Artificial Analysis scored GPT-6 Luna (max) at 37 on its Intelligence Index at launch, level with GPT-5.6 Luna (max), and at 38 on version 4.3.2 of the index when Hokai re-checked it on 19 October 2026. Its Coding Agent Index came in at 41, two points below GPT-5.6 Luna's 43. Artificial Analysis also flagged regressions of roughly 75 Elo points on GDPval-AA v2.1 and 45 Elo points on AA-Briefcase v1.1 against the predecessor, tracing the drop to weaker presentation quality and deliverables that skip rubric elements.

The model has a 1,050,000-token context window (922,000 tokens of input headroom plus up to 128,000 output tokens) and a knowledge cutoff of May 18, 2026, per OpenAI's API documentation. Input modalities are text and image; output is text only, so Luna cannot natively read or generate audio or video despite the GPT-6 branding. Reasoning effort is adjustable across six levels, none, low, medium (the default), high, xhigh and max, and OpenAI's own results show most of Luna's capability gains appear only at the max setting.

Distribution widened after launch. On 7 October 2026 OpenAI made GPT-6 Luna the model behind ChatGPT's Chat tab for Free and Go users, rolling out from 8 October, while paid plans run GPT-6 Sol there; before that, Free and Go users could reach Luna only in the desktop app. It also ships through the OpenAI API (Responses and Chat Completions), ChatGPT Work and Codex on paid plans, and GitHub Copilot across VS Code, Visual Studio, the Copilot CLI, JetBrains, Xcode, Eclipse and github.com, with GitHub confirming its own rollout the same day as OpenAI's announcement.

Through the Responses API, Luna supports web and file search, image generation, a code interpreter, hosted shell access, apply-patch, skills, computer use (billed per tool call on top of tokens; our OpenAI agents and computer-use pricing breakdown covers those fees), MCP and tool search, alongside function calling, structured outputs and streaming; the older Chat Completions endpoint only allows function calling when reasoning effort is set to none. OpenAI has not published a standalone system card for Luna: its September 22, 2026 revision to the GPT-6 Astra system card added information covering the other GPT-6 variants rather than issuing a separate document. Company-wide, API prompts and completions are excluded from OpenAI's model training by default, logs stick around for up to a month to monitor for abuse, and Zero Data Retention is available to eligible enterprise customers; it also holds SOC 2 Type II and ISO/IEC 27001 certification and can sign HIPAA business associate agreements.

Where the alternative wins: Claude Haiku 5.5, released on 7 October 2026 at an identical list price, scores 43 on Artificial Analysis's Intelligence Index and streams about 234 tokens per second, so a team that cares about answer quality or speed per call should test Haiku first. Luna's case is volume: Haiku generated roughly three times as many tokens over the same index run, which put its cost at $0.21 per task. Gemini 3.5 Flash Lite stays the natural pick for teams already on Google's API, and DeepSeek V4.1 Flash sits in the same price band for teams open to a non-US provider.

Luna fits teams that need cheap, fast inference at scale: support-ticket classification, document extraction, chat, and lightweight agent loops where OpenAI's tool ecosystem matters more than top-tier reasoning. Teams running frontier coding, research or computer-use workloads are better served by GPT-6 Sol, its 29 September successor GPT-6.1 Sol, or GPT-6 Astra, since Luna's Coding Agent Index actually fell against its own predecessor. Teams needing audio or video I/O will need a different model entirely; the context window and price will not fix a missing modality.

Pricing

OpenAI lists GPT-6 Luna at $0.10 per 1M input tokens and $0.50 per 1M output, half the $0.20 and $1.20 its predecessor charged, per OpenAI's launch post. Cached input is $0.01 and cache writes $0.125 per 1M. Batch and Flex bill at 50% of Standard, Fast mode at 2x, regional processing adds 10%, and a request with more than 272,000 input tokens is charged 2x input and 1.5x output in full, per OpenAI's model page (checked 19 October 2026).

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Wide Context Window: Most of the token budget is reserved for input, with a separate, smaller allowance for output, per OpenAI's API documentation.
• Adjustable Reasoning Effort: Six effort levels, none through max, trade speed for accuracy; OpenAI's own agentic benchmark gains mostly appear only at max effort.
• Full OpenAI Tool Stack via the Responses API: Web and file search, code interpreter, hosted shell, computer use, MCP and tool search are all available, though Chat Completions only allows function calling at the none effort level.
• Prompt Caching: Cached input is billed at a steep discount versus a fresh prompt, OpenAI's standard mechanic for repeated prompt prefixes.
• Cross-Surface Availability: Ships alongside GPT-6 Sol in OpenAI's consumer, developer and coding-assistant surfaces on day one, from the chat app to IDE plugins.

Pros

• Cuts OpenAI's own per-token price in half versus its predecessor's promotional rate, per OpenAI's September 2026 launch post.
• 126 output tokens/sec on Artificial Analysis's October 2026 run and a context window above one million tokens make it quick and roomy for a bottom-tier model.
• Ships identically across the API, ChatGPT's free and paid tiers, Codex and GitHub Copilot on day one, so teams already on OpenAI don't wait for platform support.

Cons

• Artificial Analysis's Coding Agent Index score fell slightly versus its predecessor at max effort, so it is not a pure across-the-board upgrade.
• Text and image only; there's no built-in way to process or produce audio or video, unlike some other 2026 frontier models in its price range.
• Function calling breaks on the older Chat Completions API unless reasoning effort is set to none, a trap for teams migrating existing integrations.
• At max reasoning effort Artificial Analysis measured a 74.33-second wait for the first token, against a 1.71-second median for its price tier, so max-effort calls suit batch work rather than live chat.

Benchmarks

• AA Coding Index: 41 cited: Artificial Analysis · 19 Oct 2026 — Composite of coding evaluations run by Artificial Analysis, 0 to 100.
• AA Intelligence Index: 38 cited: Artificial Analysis · 19 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $0.08/M cited: Artificial Analysis · 19 Oct 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 126 tok/s cited: Artificial Analysis · 19 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit GPT-6 Luna Official Page