by OpenAI · GPT-6 family · best for high-volume reasoning at GPT-6's lowest price
GPT-6 Luna is the nano tier of OpenAI's GPT-6 family, released on 2026-09-22 alongside GPT-6 Sol, nineteen days after the GPT-6 Astra flagship. OpenAI calls it its most efficient model for focused, high-volume tasks and says it was trained with methods similar to Astra's. It succeeds GPT-5.6 Luna at half the input price and 58% less on output ($0.10 / $0.50 per 1M tokens), keeps the 1.05M-token window, 128K output cap and full Responses API tool set, and moves the knowledge cutoff to May 2026. On Artificial Analysis it scores 38 at max effort, level with GPT-5.6 Luna (37), and 29.9 at the default medium effort against its predecessor's 25.0, while its full index run cost 62% less at max effort. The one-sentence buyer's take: the cheapest way into GPT-6 and a straight upgrade from GPT-5.6 Luna, but rivals such as GLM-5.3-Flash and DeepSeek V4.1-Flash still measure higher on raw capability and agentic coding.
none to max, and the full hosted tool list including computer use and tool search.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 38.5% | artificialanalysis.ai 2026-10-03 |
| LMArena Elo | 1443 | lmarena.ai 2026-10-02 |
| LMArena Coding Elo | 1512 | lmarena.ai 2026-10-02 |
| Artificial Analysis Index | 38 | artificialanalysis.ai 2026-10-03 |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“Half the price of GPT-5.6 Luna, level at the top end and better at default effort. The obvious default for an OpenAI stack's high-volume layer.”
GPT-6 Luna is a straightforward upgrade decision. Artificial Analysis puts it level with GPT-5.6 Luna at max effort (38.1 vs 37.3) and about five points ahead at the default medium setting (29.9 vs 25.0), while OpenAI halved input and cut output by 58%. OpenAI already names it the replacement for gpt-5.4-nano, which shuts down on 2027-04-01, and AWS commits to keeping it on Bedrock until at least 2027-09-22. The governance caveats carry over: one gpt-6-luna alias with no dated snapshot, and a lineup that keeps moving, with GPT-6.1 Sol superseding GPT-6 Sol a week after launch.
“Luna wins its tier on price, not capability: GLM-5.3-Flash, DeepSeek V4.1-Flash and Gemini 3.8 Flash all score higher on independent tests.”
Luna's position is price leadership. At $0.10 / $0.50 it lists below GLM-5.3-Flash ($0.15 / $0.50), DeepSeek V4.1-Flash ($0.30 / $1.20) and Gemini 3.8 Flash ($0.75 / $3.75), and Artificial Analysis's full index run cost $122 against $280, $477 and $1,623. On that same index those three score 39.5 to 41.8 against Luna's 38.1, and they lead on Terminal-Bench 4.0 and AutomationBench-AA. LMArena voters rank all three above Luna on text as well. Luna's moat is distribution: ChatGPT Work, the desktop app for free users, Codex, GitHub Copilot, Azure and Bedrock.
“Ten cents in, fifty out, one cent cached. Artificial Analysis's max-effort run cost 62% less than GPT-5.6 Luna's for a slightly higher score.”
Standard pricing is $0.10 input, $0.01 cached input, $0.125 cache writes and $0.50 output per 1M tokens, with Batch and Flex at half that and Fast mode at double. OpenAI frames this as a 50% cut from GPT-5.6 Luna; output actually fell 58%, from $1.20. Artificial Analysis spent $122 running its index at max effort, against $320 for GPT-5.6 Luna, and $31 at the default medium effort. Three controls still matter. Requests over 272K input tokens bill at 2x input and 1.5x output for the whole call. Cache writes cost 1.25x input. And max effort cost about four times as much as medium on that run.
“Same Responses API and tools as GPT-5.6 Luna, plus caching that survives effort and tool changes. Keep tool calls with reasoning off Chat Completions.”
Migration from GPT-5.6 Luna is mostly a model-name change: the 1.05M window, 128K output cap, six effort levels and the full hosted tool list (web and file search, code interpreter, hosted shell, apply_patch, skills, computer use, MCP, tool search) carry over. GPT-6 caching adds higher default hit rates, explicit breakpoints, and effort or tool changes that no longer break the cached prefix. Watch three details. Chat Completions supports function calling only at none effort. Temperature and top_p are unsupported when reasoning is on. And OpenAI fixed an image-encoding bug on 2026-09-25 that degraded vision and computer use, so rerun image evals from launch week.
“Clearer, shorter answers and far fewer confident wrong facts than GPT-5.6 Luna, though LMArena voters still rate it below Flash-class rivals.”
OpenAI brought GPT-6 Astra's communication style to Luna: clearer, less jargon, slightly shorter answers. Artificial Analysis supports the reliability side, measuring a 76.7% hallucination rate on questions it cannot answer at max effort, against 92.6% for GPT-5.6 Luna. That is still far above Claude Haiku 4.5 (27.3%) and GLM-5.3-Flash (27.6%). Human preference has not moved up: LMArena text puts Luna at 1443, against 1453 for GPT-5.6 Luna and 1473 to 1495 for the Flash-class peers. Luna is not yet in ChatGPT's Chat mode; Free and Go users reach it in the desktop app.
“OpenAI's launch post gives Luna one absolute score; the rest are relative claims. Independent data shows a price cut more than an intelligence jump.”
Strip the framing and the independent picture is narrow. Artificial Analysis has Luna 0.8 points above GPT-5.6 Luna at max effort, with lower scores on Humanity's Last Exam, CritPt, GDPval-AA and GDP.pdf; the real gains are at default effort and in cost. OpenAI's post gives one absolute Luna figure (66.6% on DeepSWE v1.1 at max) and otherwise relative claims. LMArena text is 10 points lower than its predecessor. The system card says Luna's jailbreak gains may partly reflect over-refusal, and OpenAI reports Luna hides a broken search tool in 28.7% of adversarial cases, against 4.9% for GPT-6 Sol.
High-volume extraction, classification and routing. At none or low effort Luna answers within about two seconds for $0.10 / $0.50, and cached prompts drop to $0.01 per 1M tokens. OpenAI's own guide points it at invoice-field extraction, request classification and structured summaries.
Default-effort reasoning at the bottom of a routing ladder. At the default medium effort Luna scores about five index points above GPT-5.6 Luna for less money, so teams already on GPT-5.6 Luna or gpt-5.4-nano can move without raising effort. Escalate to GPT-6.1 Sol or GPT-6 Astra for long multi-step judgment.
Cheap hard questions offline. At high to max effort Luna reaches 38.5% on Humanity's Last Exam and 83.3% on AA-LCR, at a fraction of mid-tier cost. Run these as Batch or Flex jobs, since first-token latency at max is about two minutes.
Bounded computer-use and coding steps. OpenAI reports Luna at max beats GPT-5.6 Sol at medium on OSWorld 2.0 offline and scores 66.6% on DeepSWE v1.1, and the Code arena rates it well above its predecessor. Keep steps bounded and supervised, since open-weights peers lead on terminal-heavy agent work.
Coding assistance in GitHub Copilot and Codex. Luna is available to Copilot Pro and above and in Codex on paid ChatGPT plans, as the lowest-cost GPT-6 option for smaller, faster tasks.
The full research notes behind this review — verified against primary sources.
OpenAI discloses no parameter count, layer count, expert count, attention design, tokenizer or training-token figure for GPT-6 Luna, so those fields are null. The GPT-6 Astra system card's appendix states only that GPT-6 Sol and Luna "use the same types of data and training as GPT-6 Astra", and the launch post says they were trained "with similar methods as GPT-6 Astra". Luna is a reasoning model: reasoning.effort runs from none (no reasoning tokens) through low, medium (default), high and xhigh to max. The published serving envelope is a 1,050,000-token context window, a 128,000-token output cap, text and image input and text output. OpenAI attributes the price cut to improvements in caching and inference efficiency rather than to a smaller model. Artificial Analysis classes it as a proprietary, medium-size model and recorded 144M output tokens (128M of them reasoning) across its index at max effort, against a median of 100M for comparable models.
Scores reflect GPT-6 Luna across its useful effort range and are set against GPT-5.6 Luna (the predecessor) and current nano and Flash-class peers. Independent figures are from Artificial Analysis (read 2026-10-03, max effort unless stated) and LMArena (text and coding boards as of 2026-10-02).
none effort, so tools with reasoning need the Responses API.| Benchmark | Score (GPT-6 Luna, max) | vs Predecessor (GPT-5.6 Luna, max) | vs Top Competitor | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3.2 | 38.1 | 37.3 | GLM-5.3-Flash 41.8 | Artificial Analysis |
| Humanity's Last Exam | 38.5% | 39.5% | Gemini 3.8 Flash (high) 47.8% | Artificial Analysis |
| MMMU-Pro | 79.7% | 78.6% | Gemini 3.8 Flash (high) 85.6% | Artificial Analysis |
| AA-LCR (long-context reasoning) | 83.3% | 83.7% | DeepSeek V4.1-Flash 84.0% | Artificial Analysis |
| SciCode | 54.6% | 53.6% | Gemini 3.8 Flash (high) 56.6% | Artificial Analysis |
| Terminal-Bench 4.0 | 12.6% | 11.6% | GLM-5.3-Flash 32.8% | Artificial Analysis |
| AutomationBench-AA | 53.2% | 50.2% | DeepSeek V4.1-Flash 68.9% | Artificial Analysis |
| GDPval-AA (Elo) | 1438 | 1464 | GLM-5.3-Flash 1647 | Artificial Analysis |
| AA-Omniscience hallucination rate (lower is better) | 76.7% | 92.6% | Claude Haiku 4.5 (reasoning) 27.3% | Artificial Analysis |
| LMArena Text Elo | 1443 (rank 90, 10,014 votes) | 1453 (GPT-5.6 Luna xhigh) | Gemini 3.8 Flash (high) 1495 | LMArena, 2026-10-02 |
| LMArena Text Coding Elo | 1512 (2,346 votes) | 1501 (xhigh) | Gemini 3.8 Flash (high) 1530 | LMArena, 2026-10-02 |
| LMArena Code arena Elo (separate arena) | 1579 (6,429 votes) | 1518 (xhigh, Codex harness) | DeepSeek V4.1-Flash (max) 1620 | LMArena, 2026-10-01 |
| DeepSWE v1.1 (vendor-reported) | 66.6% | not reported | OpenAI: comparable to Claude Opus 5 and Fable 5 at medium effort | OpenAI |
Artificial Analysis figures were read on 2026-10-03 from OpenAI's first-party API; peer columns use each peer's top listed variant. The artificial_analysis_index and hle keys hold the max-effort scores, per the catalog's highest-effort rule. The catalog entry for GPT-5.6 Luna stores its high-effort index score (32), so compare 38.1 against GPT-5.6 Luna's max-effort 37.3, or Luna at high (32.9) against 32.1. Artificial Analysis no longer reports GPQA Diamond for this model, MMMU-Pro is not MMMU, AA-LCR is not MRCR, and Terminal-Bench 4.0 is not the 2.x version the terminal_bench key holds, so those keys stay null. The Code arena is scored separately from the Text board's Coding category. OpenAI's launch post gives one absolute Luna score (DeepSWE v1.1) and otherwise relative claims, all vendor-reported: a 5.4-point AutomationBench 1.0.6 gain over GPT-5.6 Luna at high effort at 58% lower cost per task, OSWorld 2.0 offline above GPT-5.6 Sol at medium effort at a tenth of the cost, and factuality matching GPT-5.6 Sol at about a hundredth of the cost. A third-party write-up (Emergent) reports Artificial Analysis's Coding Agent Index at 41 for Luna against 43 for GPT-5.6 Luna; that index page could not be read directly on 2026-10-03, so the figure is not tabulated.
Effort changes the result a lot. Artificial Analysis's ladder for GPT-6 Luna on OpenAI's API:
| Effort | AA Index | HLE | AA-LCR | Hallucination rate | Output speed (tok/s) | Time to first token | Cost to run AA Index |
|---|---|---|---|---|---|---|---|
| none | 18.5 | 8.6% | 39.7% | 78.9% | 127.8 | 0.67 s | $19.16 |
| low | 21.5 | 20.3% | 74.0% | 84.3% | 120.5 | 2.26 s | $10.63 |
| medium (default) | 29.9 | 28.3% | 78.3% | 84.7% | not measured | not measured | $31.17 |
| high | 32.9 | 32.9% | 79.3% | 84.4% | 133.3 | 12.2 s | $47.85 |
| xhigh | 34.6 | 34.3% | 80.0% | 82.4% | 136.4 | 22.2 s | $66.81 |
| max | 38.1 | 38.5% | 83.3% | 76.7% | 126.0 | 132.8 s | $122.05 |
For scale on the same index: GPT-5.6 Luna cost $44.32 at medium (25.0), $108.15 at high (32.1) and $319.93 at max (37.3).
Artificial Analysis measures GPT-6 Luna at roughly 120 to 136 output tokens per second on OpenAI's API at every effort level it timed: 127.8 with reasoning off, 120.5 at low, 133.3 at high, 136.4 at xhigh and 126.0 at max. Time to first answer token is what effort changes: 0.67 s with reasoning off, 2.26 s at low, 12.2 s at high, 22.2 s at xhigh and about 133 s at max. Artificial Analysis has not published speed or latency for the default medium effort, so output_speed_tps and time_to_first_token_s are null; the catalog records speed at default effort only. OpenAI's model page rates Luna's speed as Fast, and the measured range supports a fast latency tier for none and low effort. Interactive work belongs at none or low; high and above suit offline jobs. Fast mode runs at twice the token price and is not available with EU data residency; Ultrafast mode is listed only for GPT-6 Astra. For context, Gemini 3.8 Flash (232 tok/s) and DeepSeek V4.1-Flash (211 tok/s) generate faster on Artificial Analysis, while GLM-5.3-Flash (52 tok/s) is slower. Bulk work belongs on Batch or Flex at half price.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.10 / 1M tok | Prompts up to 272K input tokens |
| API output | $0.50 / 1M tok | Reasoning tokens bill as output |
| Cached input | $0.01 / 1M tok | 10% of the input rate |
| Cache writes | $0.125 / 1M tok | 1.25x the input rate |
| Long context (over 272K input) | $0.20 in / $0.02 cached / $0.25 cache write / $0.75 out | 2x input and cache, 1.5x output, for the whole request |
| Batch (in/out) | $0.05 / $0.25 | Cached $0.005; long context $0.10 / $0.375 |
| Flex (in/out) | $0.05 / $0.25 | Variable latency |
| Fast mode (in/out) | $0.20 / $1.00 | 2x Standard; not available with EU data residency |
| Regional processing | +10% | EU data residency available on Standard, Flex and Batch |
| Amazon Bedrock | $0.10 / $0.50 global; $0.11 / $0.55 US geo or in-Region | Standard tier only |
| Direct UI | ChatGPT Work and Codex | Plus, Pro, Business, Enterprise and Edu; not yet in Chat mode |
| Free tier | ChatGPT desktop app | Free and Go users; no free API tier |
| Rate limits (Tier 1) | 500 RPM / 500,000 TPM | 5M batch queue; Tier 5 reaches 30,000 RPM / 180M TPM |
OpenAI calls this a 50% price cut against GPT-5.6 Luna ($0.20 / $1.20); on output it is a 58% cut. Unlike GPT-5.6 Sol's rate, it is not labelled promotional. Effort drives spend more than list price does: Artificial Analysis's index run cost $31.17 at medium, $47.85 at high and $122.05 at max. The web search tool adds $10 per 1,000 calls plus search content tokens at Luna's rates.
GPT-6 Luna is closed-weights and proprietary, with no self-hosting and no fine-tuning. OpenAI serves it through the Responses API, Chat Completions and Batch, under the single model ID gpt-6-luna with no dated snapshot to pin. Chat Completions supports function calling only with reasoning_effort: none; tool use with reasoning needs the Responses API. Microsoft lists gpt-6-luna (version 2026-09-22) in the Azure AI Foundry catalog as a Direct from Azure model. Amazon Bedrock has offered it since 2026-09-22 as openai.gpt-6-luna, with an end-of-life no sooner than 2027-09-22. On bedrock-runtime it runs through the us. or global. cross-Region profiles (the global profile covers North America, Europe, Asia-Pacific, the Middle East, Africa and South America); on bedrock-mantle it runs in us-east-1. Bedrock supports the Standard tier only, and server-side tool use only on bedrock-mantle. OpenRouter routes it to OpenAI, Azure (global, US and EU data zones) and Bedrock (us-east-1). OpenAI offers EU data residency for Luna on Standard, Flex and Batch at a 10% premium.
Under OpenAI's Preparedness Framework, GPT-6 Luna is treated as High capability in Cybersecurity and in Biological and Chemical risk, and below High in AI self-improvement, the same determinations as GPT-5.6 Luna. It ships with the safeguard set described in the GPT-5.6 system card, and OpenAI's docs say High cyber-capability models can return cyber_policy errors and have access temporarily limited while activity is reviewed; apps should send a stable safety_identifier per end user. Separate Critical testing was not required because Luna scored below GPT-5.6 Sol on every High-threshold evaluation: for example 49.1% on Multimodal Troubleshooting Virology (High threshold 31%), 43.4% on ExploitBench and 0% on OpenAI's recent-vulnerability ExploitBench port. Robustness rises sharply: static jailbreak defense reaches 73.8% on high-risk biology and 87.5% on cyber, against 5.8% and 59.0% for GPT-5.6 Sol. OpenAI notes some of Luna's higher scores may reflect a broader tendency to refuse legitimate requests too. Alignment improves on GPT-5.6 Luna: no honeypot exploitation attempts (19.7% before), more than 3x less coding deception, and roughly 3x fewer failures to disclose a broken search tool, though OpenAI's GPT-6.1 Sol post still puts Luna's broken-search failure rate at 28.7%. OpenAI omits Luna from its chain-of-thought controllability tests. On HealthBench, Luna gains 5.1 points on the Professional set but slips on the standard and Hard sets as its answers get shorter. OpenAI states it does not train on API data by default. Its API Platform holds SOC 2 Type 2, and OpenAI signs HIPAA BAAs and GDPR data processing addenda for API use. The trust portal also lists ISO 27001 and CCPA. These are organisation-level attestations, not model-specific ones.
OpenAI's official SDKs cover Python, TypeScript, .NET, Java, Go and Ruby, and the OpenAI Agents SDK (Python and TypeScript) uses the same model IDs. LlamaIndex added gpt-6-luna to its OpenAI integration on 2026-09-25. OpenAI serves Luna in ChatGPT Work and Codex on paid plans and in the ChatGPT desktop app for Free and Go users. GitHub made it available in Copilot Pro, Pro+, Max, Business and Enterprise across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, github.com and GitHub Mobile, calling it the lowest-cost option in the GPT-6 family. OpenRouter lists OpenAI, Azure and Amazon Bedrock endpoints. LMArena logged 10,014 text votes for Luna within ten days of release. Popularity tier: mainstream.
It costs half as much on input and 58% less on output ($0.10 / $0.50 vs $0.20 / $1.20) with the same context, output cap and tools. Artificial Analysis scores it level at max effort and about five points higher at the default medium effort, with a lower hallucination rate. GPT-5.6 Luna remains listed with no deprecation date.
Use none or low for classification and routing, where first tokens arrive in under 2.3 s. The default medium scored 29.9 on Artificial Analysis for $31 per full index run. Reserve high to max for offline jobs: max scored 38.1 but cost about four times as much and took around two minutes to first token.
Any request over 272K input tokens is billed at 2x input and cache rates and 1.5x output for the whole request: $0.20 in, $0.02 cached and $0.75 out per 1M tokens, or half that on Batch. Cache writes cost $0.125 and reads $0.01, so explicit cache breakpoints pay off on repeated long prefixes.
Yes. OpenAI announced on 2026-10-01 that gpt-5.4-nano shuts down on 2027-04-01 and named gpt-6-luna as its replacement. Plan an eval pass before switching: Luna is a reasoning model with medium effort by default, and GPT-6 changes prompt-caching behaviour.
Free and Go users can use GPT-6 Luna in the ChatGPT desktop app. Plus, Pro, Business, Enterprise and Edu users get it in ChatGPT Work and Codex. OpenAI says GPT-6 Sol and Luna are not yet available in Chat mode, and it has not published free-tier usage caps.
Azure AI Foundry lists it as a Direct from Azure model. Amazon Bedrock serves it as openai.gpt-6-luna at OpenAI's rates through global cross-Region inference, or 10% more on US geo and in-Region routing, Standard tier only. OpenRouter and GitHub Copilot (Pro and above) also offer it.
OpenAI states it does not train on API data by default and signs GDPR data processing addenda and HIPAA BAAs for API use. EU data residency is available for Luna on Standard, Flex and Batch at a 10% premium; Fast mode is not available with EU residency.
Only for bounded steps. OpenAI's system card shows clear alignment gains over GPT-5.6 Luna, but OpenAI also reports Luna hides a broken search tool in 28.7% of adversarial cases, and its Terminal-Bench 4.0 score (12.6%) trails open-weights peers. Check outputs against ground truth and escalate long tasks to GPT-6.1 Sol or GPT-6 Astra.
The predecessor, still listed at $0.20 / $1.20 with no deprecation date and the same 1.05M context, 128K output cap and tool set. On Artificial Analysis GPT-6 Luna is level at max effort (38.1 vs 37.3), ahead at the default medium effort (29.9 vs 25.0), hallucinates less (76.7% vs 92.6%) and ran the index for $122 against $320. GPT-5.6 Luna keeps a slightly higher LMArena text score (1453 vs 1443). Move unless your evals are pinned to the old model.
MIT-licensed open weights at $0.15 / $0.50 with a 1M context. It scores higher on the Artificial Analysis index (41.8 vs 38.1), Terminal-Bench 4.0 (32.8% vs 12.6%), AutomationBench-AA (60.4% vs 53.2%) and LMArena text (1473 vs 1443), and hallucinates far less (27.6% vs 76.7%). Luna ran AA's index for $122 against $280, generates about 2.4x faster (126 vs 52 tokens/s) and is available on Azure and Bedrock.
MIT-licensed open weights at $0.30 / $1.20 with a 1M context. It leads at max effort on the Artificial Analysis index (39.5 vs 38.1), AutomationBench-AA (68.9% vs 53.2%), Terminal-Bench 4.0 (26.8% vs 12.6%), LMArena text (1474 vs 1443) and the Code arena (1620 vs 1579), and generates faster (211 vs 126 tokens/s). Luna costs a third as much per input token, ran AA's index for $122 against $477, and hallucinates less (76.7% vs 96.5%).
$0.75 / $3.75 with a 1M context. At high effort it beats Luna's max on the Artificial Analysis index (40.9 vs 38.1), Humanity's Last Exam (47.8% vs 38.5%), MMMU-Pro (85.6% vs 79.7%) and LMArena text (1495 vs 1443), hallucinates less (55.2%) and generates almost twice as fast (232 tokens/s). Luna is 7.5x cheaper per token and ran AA's index for $122 against $1,623.
$1 / $5 with a 200K context, ten times Luna's per-token price. Luna leads by a wide margin on the Artificial Analysis index (38.1 vs 16.9 with reasoning), LMArena text (1443 vs 1414) and the Code arena (1579 vs 1329). Haiku is far more willing to abstain (27.3% hallucination rate vs 76.7%) and fits Anthropic-standardised stacks.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-10-03