by OpenAI · GPT-5 family · best for high-volume reasoning worker at nano-tier price
GPT-5.6 Luna is the nano tier of OpenAI's GPT-5.6 family, released to the API on 2026-07-09 alongside GPT-5.6 Sol and GPT-5.6 Terra. OpenAI describes it as the GPT-5.6 model for cost-sensitive, high-volume workloads and names it the replacement for gpt-5-nano and gpt-4.1-nano. It costs the same $0.20 per 1M input tokens as GPT-5.4 nano, but brings a 1.05M-token context window, the full Responses API tool surface including computer use, and a reasoning dial that runs from none to max. On Artificial Analysis it scores 25 on the Intelligence Index at its default medium effort and 32 at high effort, ahead of GPT-5.4 nano, GPT-5.4 mini, Gemini 3.5 Flash-Lite and Claude Haiku 4.5 at any setting. The one-sentence buyer's take: the default cheap worker for OpenAI-stack teams, and a real reasoning model at nano prices, as long as answers that depend on recalled facts are grounded with retrieval.
none, low, medium (default), high, xhigh and max. Pro mode (reasoning.mode: "pro") works on Luna and is billed at Luna's standard token rates.reasoning.context: all_turns), and images accepted at their original dimensions.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 33.4% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| LMArena Elo | 1452 | lmarena.ai 2026-09-13T00:00:00.000Z |
| GPQA Diamond | 89.2% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| LMArena Coding Elo | 1498 | lmarena.ai 2026-09-13T00:00:00.000Z |
| Artificial Analysis Index | 32 | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The cheapest OpenAI model now reasons like last year's mid tier. Standardise the worker layer on it and escalate by model name, not vendor.”
Luna changes the default choice for the bottom of an OpenAI stack. At GPT-5.4 nano's price it measures above GPT-5.4 mini on the Artificial Analysis Intelligence Index, adds a 1.05M window and the full tool surface, and is OpenAI's named replacement for gpt-5-nano and gpt-4.1-nano. Vendor risk is modest: it is available on Azure and Amazon Bedrock as well as OpenAI's API, so procurement has options. The caveats are governance ones. There is only one alias with no dated snapshot, and the price moved 80% within three weeks of launch, so lock regression tests into the rollout plan.
“Luna resets the nano floor: Flash-Lite and Haiku now look expensive for what they measure, and only open-weights DeepSeek beats it.”
In the sub-$0.50 proprietary segment Luna leads on measured intelligence. Its index score of 32 at high compares with 22 for Gemini 3.5 Flash-Lite and 17 for Claude Haiku 4.5, and its output price is lower than either. That puts pressure on Google and Anthropic in the high-volume layer. The competitive threat comes from open weights. DeepSeek V4.1-Flash scores higher at max effort, costs less off-peak ($0.15 / $0.60) and ships under MIT. Mistral Small 4 is cheaper again under Apache-2.0 but measures far lower. Luna's moat is distribution: ChatGPT free users, Replit, Azure and Bedrock, plus the most complete tool surface in its tier.
“Twenty cents in, a dollar twenty out, two cents cached. Luna buys GPT-5.4-mini-class answers at nano rates if you control effort and length.”
List pricing is $0.20 / $1.20, with cached input at $0.02, Batch and Flex at $0.10 / $0.60, and Fast mode at double the standard rate. The value case is strong. On Artificial Analysis, Luna at high effort scored 32 for $108 across the whole index, while GPT-5.4 nano at xhigh scored 21 for $331. Three cost traps need controls. Prompts over 272K tokens double input pricing for the whole call. GPT-5.6 bills cache writes at 1.25x input. Max effort roughly triples spend over high. Cap effort per route and track cache-write tokens to keep the budget predictable.
“Same Responses API, now with computer use, tool search and a none-to-max effort dial. Set effort per call and it covers most worker jobs.”
For builders, Luna removes GPT-5.4 nano's main gaps. Computer use and tool search are supported, and the GPT-5.6 additions (Programmatic Tool Calling, explicit cache breakpoints, persisted reasoning) work on it. The 128K output cap and 1.05M window are generous, and structured outputs and MCP behave as on the rest of the family. The practical work is tuning. Default medium effort may be slower than a nano-tier user expects, reasoning persistence now defaults to all_turns, and cache writes cost extra, so migrations from GPT-5.4 need a quick pass over effort, caching and conversation-state settings.
“Fast and sharp at low effort, but it states made-up facts with full confidence. Turn on web search before trusting it on names, dates or figures.”
Luna now runs ChatGPT's free-tier everyday chats, and it holds up. LMArena voters place it level with Gemini 3.5 Flash-Lite and above Claude Haiku 4.5. Responses are concise by default, and with reasoning off the first token arrives in under a second. The weakness shows on lookups. Artificial Analysis finds Luna answers rather than abstains on questions it cannot know, so obscure facts come back confidently wrong far more often than with Haiku or Flash-Lite. Occasional multi-second pauses from the safety classifiers are noticeable on security or biology topics.
“Strong index scores hide a 92% hallucination rate, no SWE-bench figure, and one unpinnable alias. Test it on your own tasks before swapping.”
Independent coverage exists but is narrow: Artificial Analysis and LMArena, about ten weeks after launch. There is no public SWE-bench Verified, Aider or tau-bench result, and OpenAI's launch-post figures are not independently reproduced. The headline gains depend on effort. At the default medium setting Luna scores 25, a solid step over GPT-5.4 nano but not the 32 to 37 quoted for high and max, which cost more and take far longer. AA-Omniscience shows a 92% hallucination rate, far worse than Haiku or Flash-Lite. Pricing fell 80% three weeks after launch, and the single alias can change without notice.
none, frontier-adjacent answers at high to max, on one model ID.High-volume extraction, classification and routing. At none or low effort Luna answers in under two seconds for $0.20 / $1.20, and prompt caching drops repeated instructions to $0.02 per 1M tokens. It is the natural bottom tier of an OpenAI routing ladder (Luna, then Terra, then Sol).
Cheap reasoning on hard single questions. At high effort Luna reaches 89% on GPQA Diamond and scores above GPT-5.4 mini at a quarter of mini's list price. Teams that previously paid mini or base-tier prices for analysis, grading or planning steps can test Luna at high first.
Long-document work under 272K tokens. Contracts, transcripts and codebases up to 272K tokens bill at base rates, and AA-LCR long-context reasoning (80%) is strong for the tier. Beyond 272K the window still works, but the surcharge applies.
Tool-using worker agents. With computer use, tool search, hosted shell and Programmatic Tool Calling, Luna can run bounded agent steps such as data gathering, form filling and batch tool calls. It should hand off to Terra or Sol when a step needs sustained multi-step judgment.
Free-tier consumer products. OpenAI runs ChatGPT's free everyday chats on Luna, and Replit built its Free Mode on it. That is a signal it holds up for cost-capped, user-facing chat.
The full research notes behind this review — verified against primary sources.
OpenAI discloses no parameter count, layer count, attention design or training-data scale for Luna, so those fields are null. It is a reasoning model with text and image input and text output, a 1.05M context window and a 128K output cap. The only structural signals OpenAI gives are the family framing ("roughly corresponds to the nano model tier used in earlier GPT-5 families") and the GPT-5.6-wide claim of better token efficiency, meaning flagship-style results with fewer output tokens. Artificial Analysis's run supports the efficiency claim at mid settings: Luna at high effort used 50M output tokens across the Intelligence Index, below the 90M median for its class.
Scores reflect Luna at its useful range (medium to high effort) and are set against GPT-5.4 nano (the predecessor) and the current nano and lite peers.
text.verbosity controls length.| Benchmark | Score (Luna, high effort) | vs Predecessor (GPT-5.4 nano, xhigh) | vs Gemini 3.5 Flash-Lite | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3.2 | 32 | 21 | 22 | Artificial Analysis |
| GPQA Diamond | 89.2% | 81.7% | 83.8% | Artificial Analysis |
| Humanity's Last Exam | 33.4% | 28.3% | 18.8% | Artificial Analysis |
| MMMU-Pro | 77.6% | 65.4% | 79.0% | Artificial Analysis |
| AA-LCR (long-context reasoning) | 80.3% | 76.7% | 76.0% | Artificial Analysis |
| SciCode | 51.6% | 47.2% | 41.3% | Artificial Analysis |
| AA-Omniscience Index (higher is better) | -12 | -29 | +5 | Artificial Analysis |
| LMArena Text Elo (Luna xhigh) | 1452 | 1402 (nano high) | 1456 | LMArena |
| LMArena Text Coding Elo (Luna xhigh) | 1498 | 1460 (nano high) | 1504 | LMArena |
| LMArena Code/WebDev arena Elo (separate arena, Luna xhigh, Codex harness) | 1519 | not listed | 1447 | LMArena |
All scores are independent measurements: Artificial Analysis figures read on 2026-09-21, LMArena text and its Coding category as of 2026-09-13 (28,547 text votes for Luna), and the separate Code/WebDev arena as of 2026-09-11. The Code/WebDev arena is scored apart from the Text board's Coding category, so its Elo figures do not compare with the Text Coding Elo. OpenAI's own launch-post figures are not reproduced here. MMMU-Pro is not the same test as MMMU, and AA-LCR is not MRCR, so those front-matter keys stay null.
Effort changes the result a lot. The same Artificial Analysis dataset gives this ladder for Luna, measured on OpenAI's API:
| Effort | AA Index | GPQA Diamond | HLE | Output speed (tok/s) | Time to first token | Cost to run AA Index |
|---|---|---|---|---|---|---|
| none | 15.5 | 64.5% | 7.2% | 153 | 0.83 s | $22.97 |
| low | 21.0 | 83.5% | 19.8% | 148 | 1.77 s | $27.02 |
| medium (default) | 25.0 | 85.9% | 25.8% | 151 | 2.63 s | $44.32 |
| high | 32.1 | 89.2% | 33.4% | 150 | 9.78 s | $108.15 |
| xhigh | 34.6 | 89.5% | 37.0% | 165 | 33.4 s | $179.33 |
| max | 37.3 | 91.1% | 39.5% | 162 | 126 s | $319.93 |
For scale: GPT-5.4 nano at xhigh scored 20.7 for $331.24, and Claude Haiku 4.5 with reasoning scored 16.9 for $524.25. Luna at high effort beats both for about a third of the cost or less.
Artificial Analysis measures Luna at about 150 output tokens per second on OpenAI's API at every effort level. What changes with effort is time to first token: 0.83 s with reasoning off, 1.77 s at low, 2.63 s at the default medium, 9.78 s at high, and around two minutes at max. The front-matter records the medium-effort figures (151 tok/s, 2.63 s), so latency_tier is fast for default use. Chat-style and routing work should run at none or low. Latency-sensitive long-context work can use Fast mode, which OpenAI says runs up to 2.5x faster than Standard at twice the token price. Separately, the GPT-5.6 safety classifiers can pause a stream for several seconds while they review output, which shows up as occasional stalls rather than slower average speed. Bulk work belongs on Batch or Flex at half price.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.20 / 1M tok | prompts up to 272K tokens |
| API output | $1.20 / 1M tok | reasoning tokens bill as output |
| Cached input | $0.02 / 1M tok | 90% discount |
| Cache writes | $0.25 / 1M tok | 1.25x input; GPT-5.6 bills cache writes |
| Long context (prompt over 272K) | $0.40 in / $0.04 cached / $1.80 out | 2x input, 1.5x output for the whole request |
| Batch (in/out) | $0.10 / $0.60 | cached $0.01; long context $0.20 / $0.90 |
| Flex (in/out) | $0.10 / $0.60 | variable latency |
| Fast mode (in/out) | $0.40 / $2.40 | up to 2.5x faster than Standard |
| Regional processing | +10% | data-residency endpoints for eligible models released on or after 2026-03-05 |
| Direct UI | ChatGPT Free | OpenAI runs free-tier everyday chats on Luna |
| Free tier (API) | none | |
| Rate limits (Tier 1) | 500 RPM / 500,000 TPM | 5M batch queue; Tier 5 reaches 30,000 RPM / 180M TPM |
Price history matters for budgeting: Luna launched on 2026-07-09 and was cut 80% on 2026-07-30. OpenAI did not describe the cut as promotional, unlike the time-limited GPT-5.6 Sol pricing. Pro mode and max effort bill at the same per-token rates but use many more tokens. At max effort, the Artificial Analysis run cost about three times as much as the high-effort run.
Luna is API-only: the Responses API, Chat Completions and Batch are supported. Realtime, Assistants and fine-tuning are not. Weights are closed, the license is proprietary, and self-hosting is not possible. OpenAI lists a single model ID, gpt-5.6-luna, with no dated snapshot to pin. Beyond OpenAI's own endpoint, OpenRouter routes Luna to Azure (global, US and EU data zones) and Amazon Bedrock (us-east-1). It lists the Azure US/EU and Bedrock endpoints at a 10% premium over OpenAI's global rate, the same uplift OpenAI charges for regional processing. Promotion to GPT-5.6 Terra or Sol is a model-name change with the same tools and parameters.
Luna falls under OpenAI's Preparedness Framework. The GPT-5.6 family runs real-time cyber and biology misuse classifiers on output as it is generated. These can refuse or block requests, and they sometimes pause a stream for several seconds while output is reviewed. OpenAI says they occasionally intervene on legitimate dual-use work such as vulnerability research. Apps that serve end users should send a stable safety_identifier with each request. API data is not used for training by default. Zero Data Retention is available to eligible customers, and Programmatic Tool Calling is ZDR-compatible. No Luna-specific compliance attestation list is published on the model page; buyers should confirm SOC 2, HIPAA and GDPR scope through OpenAI's organisation-level trust documentation.
First-party SDKs cover Python, TypeScript, Java, Go and .NET, and the OpenAI Agents SDK and Codex use the same model IDs. Framework support through the OpenAI API extends to LangChain, LlamaIndex and the Vercel AI SDK. Luna runs ChatGPT's free-tier everyday chats and Replit's Free Mode. It is listed on OpenRouter with OpenAI, Azure and Amazon Bedrock endpoints. It had 28,547 LMArena text votes by 2026-09-13, which puts it in mainstream use within ten weeks of release. Popularity tier: mainstream.
Input costs the same $0.20 per 1M tokens and output is slightly cheaper ($1.20 vs $1.25). Luna adds a 1.05M context window (vs 400K), computer use and tool search, a newer 2026-02 knowledge cutoff, and measures about five index points higher at default effort on Artificial Analysis. GPT-5.4 nano remains GA with no deprecation date, so there is no forced migration.
Start at none or low for classification and routing (under 2 s to first token). Use the default medium for general work. Test high for analysis, where Luna's index score jumps from 25 to 32 for about 2.4x the evaluation cost. Reserve xhigh and max for offline jobs, because time to first token reaches roughly 30 s and two minutes.
Any request with more than 272K input tokens is billed at 2x input and 1.5x output for the whole request: $0.40 / $1.80 per 1M tokens, or $0.20 / $0.90 on Batch. Cache writes cost $0.25 per 1M tokens and cache reads $0.02, so explicit cache breakpoints pay off on repeated long prefixes.
Yes. The gpt-5-nano snapshot shuts down on 2026-12-11 and gpt-4.1-nano on 2026-10-23, and OpenAI names Luna as the replacement for both. Budget for the change: Luna's $0.20 / $1.20 is 4x gpt-5-nano's $0.05 / $0.40 on input and 3x on output.
It supports the full Responses API tool set, including computer use, tool search, hosted shell, MCP and Programmatic Tool Calling. Independent agentic scores are good for the tier but well below DeepSeek V4.1-Flash. Use Luna for bounded worker steps and escalate long multi-step tasks to Terra or Sol.
API data is not used for training by default, and Zero Data Retention is available to eligible customers. Regional processing endpoints carry a 10% uplift. Luna is also served from Azure US and EU data zones and from Amazon Bedrock.
DeepSeek V4.1-Flash (MIT) matches or beats Luna on independent scores and can be self-hosted, but needs large-GPU infrastructure. Mistral Small 4 (Apache-2.0, $0.15 / $0.60, 256K context) is cheaper and easy to host but scores far lower (11.3 on the Artificial Analysis index with reasoning).
The current rate followed an 80% cut on 2026-07-30, three weeks after launch. OpenAI did not label it promotional, unlike GPT-5.6 Sol's time-limited pricing, but there is no contractual price lock outside enterprise agreements.
$0.30 / $2.50 with a 1,048,576-token input limit and 65,536-token output cap, plus native audio, video and PDF input and Search grounding. It ties Luna on LMArena text (1456 vs 1452), edges it on MMMU-Pro (79.0% vs 77.6%) and hallucinates far less (34% vs 92% on AA-Omniscience). It trails on the Artificial Analysis Intelligence Index (22 vs 32 at high effort) and costs about twice as much per output token. Pick Flash-Lite for modalities and factual restraint, Luna for reasoning per dollar.
$1 / $5 (five times Luna's input price), 200K context, 64K output, knowledge cutoff February 2025. It scores lower on the Artificial Analysis index (16.9 with reasoning), LMArena text (1415) and LMArena's separate Code/WebDev arena (1329). It is far more willing to abstain (27% hallucination rate) and fits Anthropic-standardised stacks. Luna wins on price, context, recency and measured capability.
MIT-licensed open weights (552B MoE), $0.30 / $1.20 at peak and $0.15 / $0.60 off-peak, 1M context and a 384K output cap. It scores above Luna at max effort on the Artificial Analysis index (39.5 vs 37.3) and on agentic work (AutomationBench-AA 68.9% vs 50.2%), and it can be self-hosted. Luna wins on US-hosted cloud availability (Azure, Bedrock), the OpenAI tool surface and predictable single-rate pricing.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21