GPT-6 Luna
GPT-6 Luna is OpenAI's fast, cost-efficient multimodal variant in the GPT-6 series, positioned below GPT-6 Sol. It offers a 1,050,000-token context window, image input, a May 18, 2026 knowledge cutoff, and selectable reasoning effort (none→max) at an aggressively low $0.10/$0.50 per 1M tokens — making it a high-volume, latency-sensitive workhorse rather than a frontier reasoning flagship.
Empirical Evaluation Results
Architectural Profile & Capabilities
GPT-6 Luna (`openai/gpt-6-luna`) is the lower-tier, cost-optimized member of OpenAI's GPT-6 family, announced alongside GPT-6 Sol on September 22, 2026 (OpenRouter record; OpenAI announcement). It is intended for high-volume, latency-sensitive workloads — chat, classification, structured extraction, intent triage, and lightweight agentic pipelines — and is priced roughly 20x below GPT-6.1 Sol on both input and output. Confirmed product-level specifications include a 1,050,000-token context window, a 128,000-token maximum output, a May 18, 2026 knowledge cutoff, and multimodal (text + image) input. It exposes six reasoning-effort levels (none, low, medium, high, xhigh, max); the Benchmark Heaven entry corresponds to the `max` reasoning configuration (`gpt-6-luna::max`). Empirical throughput measured by Benchmark Heaven is ~128 tokens/sec. OpenAI has NOT publicly disclosed the underlying architecture — no verified confirmation of dense vs. sparse Mixture-of-Experts, no parameter counts, and no confirmed attention mechanism (MLA/GQA). Claims that Luna is MoE or uses a specific attention scheme remain UNVERIFIED and are therefore not asserted here. Beyond the immutable Benchmark Heaven ground-truth set, this investigation identified a small number of additional exact-variant scores explicitly published for GPT-6 Luna (notably GPQA Diamond = 90.5%, corroborated independently by two leaderboards), which has been appended. All other requested benchmarks (SWE-bench Verified, AIME 2026 as a standalone, MATH-500, LiveCodeBench, BrowseComp, OSWorld Verified, TAU-2, MRCR v2, etc.) could not be verified for the exact Luna variant and are omitted per the anti-proxy mandate. Note that many leaderboards (e.g., Artificial Analysis Harvey LAB-AA, LLM Stats) list Luna in multiple reasoning-effort columns (Max, xhigh, Non-reasoning) — scores must be attributed to the correct effort level and never blended with GPT-6 Sol, Astra, or GPT-5.6 Luna.
OpenAI has not publicly documented GPT-6 Luna's internal architecture: total/active parameter counts, expert routing, and attention mechanism (MLA vs GQA) are all undisclosed. Verified surface facts: 1,050,000-token context window, 128,000-token max output, image+text modality, knowledge cutoff May 18, 2026, and six selectable reasoning-effort levels (none, low, medium, high, xhigh, max). The 'max' effort level implies substantial test-time reasoning-token generation, but no numerical reasoning-token budget is published. Because neither the vendor nor the technical report confirms a Mixture-of-Experts design, isMoe is set to false (unverified) rather than guessed. Treat all architecture claims beyond the above as unverified.
Recommended Workloads & Primary Use Cases
- •Architecture is undisclosed: no public confirmation of dense vs. MoE, parameter counts, or attention type — capacity planning must be empirical, not structural.
- •Reasoning-effort sensitivity: scores are reported at specific effort tiers (max, xhigh, non-reasoning); the ~128 tps throughput reflects max-effort settings, and higher effort raises tokens and latency while lower effort degrades hard reasoning.
- •Positioned below Sol/Astra: it is deliberately not a frontier reasoning model, so very long-horizon or SOTA-competitive tasks should route to a larger sibling rather than over-driving Luna's effort levels.
- •Long-context economics and TTFT: while the window is ~1.05M tokens, prefill latency and cost scale sharply at extreme context lengths; effective reliability will degrade well before the nominal limit for needle-in-haystack and multi-hop tasks.
- •Tool-calling/agentic reliability is uneven by domain — agentic gains (AutomationBench, SWE-style) far exceed weak frontier-agent marks (e.g., very low Harvey LAB-AA legal-agent scores), so agent scaffolds need per-task validation.
- •Reasoning-effort inflation risk: max-effort reasoning can consume budget without proportional score gains; tuning effort per task is required to preserve Luna's cost advantage.
Calls to GPT-6 Luna are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by OpenAI.
More Models from OpenAI
Compare other engines available in this laboratory.
OpenAI's flagship September 2026 frontier model with a 1.05M context window, MoE architecture, and SOTA reasoning across coding, formal mathematics, and long-horizon agentic automation.
GPT-6.1 Sol (openai/gpt-6.1-sol) is OpenAI's September 29, 2026 mid-tier reasoning model, an upgrade to GPT-6 Sol positioned one step below the flagship GPT-6 Astra. It targets agentic coding, computer use, and document-heavy professional automation at roughly one-fifth of Astra's per-token cost, with a 1.05M-token context window, text+image input, and a measured output throughput of ~53 tokens/sec.
GPT-6 Sol is OpenAI's mid-tier, cost-efficient high-end GPT-6 model (launched 22 Sept 2026, ~three weeks after flagship GPT-6 Astra), exposing a 1,050,000-token context window, 128K max output, image input, selectable reasoning effort, and hidden test-time reasoning tokens at $2/$10 per 1M tokens — positioned between GPT-6 Astra (quality) and GPT-6 Luna (throughput/price).