Benchmarks
More
OverviewBenchmaxxingCompareChartsCost vs CapabilityProviders per ModelProvider explorerGatewaysEU & SovereignJev-class modelsImage JevBenchAudioJevBenchSubmit a modelAbout
← All models

GPT-6 Luna (Max)

OpenAI · released 2026-09-22 · 9 offers

Output 120 tokens/sFirst token 90 sContext 1M tokens

Top 3 cheapest providers (Adjusted $/task)

The same list price can give a different adjusted $/task (caching, token efficiency) — click a price for its inputs.

Within the active global provider, residency and confidentiality filters.

#ProviderAdjusted $/task
1Azurevia OpenRouter — Price via OpenRouter: OpenRouter's price for Azure's endpoint (https://openrouter.ai/openai/gpt-6-luna/providers), collected 2026-11-05.
2Amazon Bedrockvia OpenRouter — Price via OpenRouter: OpenRouter's price for Amazon Bedrock's endpoint (https://openrouter.ai/openai/gpt-6-luna/providers), collected 2026-11-05.
3T-Systems LLM Hubdirect — Direct price: T-Systems LLM Hub's own price list (https://docs.llmhub.t-systems.net/models/llms/), collected 2026-11-05.

Composite

5 of 7 inputs · 4 from the model family6 radar axes: DesignArena's two boards share one

53.2

AA Coding v4.xCoding Agent v1.4AA IntelligenceAA AgenticEpoch ECISoftware ECIDesignArenaAA Intelligence: percentile 88Epoch ECI: percentile 59Software ECI: percentile 58DesignArena: percentile 36

AA Coding v4.x —Coding Agent v1.4 —AA Intelligence 27.9AA Agentic —Epoch ECI 156.7Software ECI 158.3DesignArena 1161/1152

Radar: percentile among all models measured on each input; a gap means not measured.

Benchmark sheet

52 of 319 registered benchmark versions · bars show the percentile among all models measured on each benchmark.

12 of 53 values are OpenAI's own claims (†), not independent measurements; a matching independent result replaces a claim as soon as one exists.

Composite attachments (used in the score, not counted as exact benchmarks):
  • Epoch ECI · attached
  • Software ECI · attached
  • DesignArena Web Apps (agentic) · attached
  • DesignArena Full-Stack · attached
Compare this model →

Agentic

  • Terminal-Bench v4.0 (AA, upstream timeouts)2912.6%

    Terminal-Bench v4.0 (AA, upstream timeouts) 4.0-upstream-timeouts → · Published board — Tests terminal-based work on the 66-task release using AA's mini-swe-agent harness with the tasks' own timeouts.

  • Finance Agent v2 (Vals Index v2)2749.9%

    Finance Agent v2 (Vals Index v2) v2 → · Published board — Multi-step financial reasoning tasks.

  • HLAB — Harvey's Legal Agent Benchmark (Vals Index v2)262.9%

    HLAB — Harvey's Legal Agent Benchmark (Vals Index v2) v2 → · Published board — Long-horizon legal work product creation. The index column is Harvey's Legal Agent Benchmark's own published standalone score, reported as accuracy — not comparable with Artificial Analysis' Harvey LAB-AA row.

  • Terminal-Bench 2.1 (Vals Index v2)4973.0%

    Terminal-Bench 2.1 (Vals Index v2) v2 → · Published board — Command-line interface problem solving; the page states that each index column is that benchmark's own published standalone score, so the runner is not asserted here.

  • Vals Index v2 (Vals AI)4958.5%

    Vals Index v2 (Vals AI) v2 → · Published board — GDP-weighted average of agentic model performance across finance, coding and legal tasks.

  • AutomationBench 1.0.6developer's claim20.7%† self-reported by the developer

    AutomationBench 1.0.6 v1.0.6 → · Vendor-reported by OpenAI — AutomationBench 1.0.6 score OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • Agents' Last Exam V1developer's claim50.9%† self-reported by the developer

    Agents' Last Exam V1 v1 → · Vendor-reported by OpenAI — Agents' Last Exam V1 score OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • OSWorld 2.0 offline (v2026.08.08 release)developer's claim52.7%† self-reported by the developer

    OSWorld 2.0 offline (v2026.08.08 release) v2026.08.08 → · Vendor-reported by OpenAI — OSWorld 2.0 offline partial reward OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • Tax Agent Bench2258.9%

    Tax Agent Bench published 2026-09-23 → · Published board — Tax Agent Bench results published by Vals AI.

Coding

  • Artificial Analysis Coding Agent Index v2.0031.1%

    Artificial Analysis Coding Agent Index v2.0 v2.0 → · Codex — Measures coding-agent systems on FrontierCode v1.1 Main, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA.

  • SciCode (AA subproblems) v1.0.17654.6%

    SciCode (AA subproblems) v1.0.1 → · Published board — Tests scientific Python programming with scientist-annotated background information.

  • FrontierCode 1.1 Main (Cognition)developer's claim42.4%† self-reported by the developer

    FrontierCode 1.1 Main (Cognition) v1.1 → · Codex — Whether a maintainer would merge the agent's pull request, on tasks crafted by open-source maintainers and graded with tests, rubrics and verifiers.

  • Code Migration (Vals Index v2 subset)6544.4%

    Code Migration (Vals Index v2 subset) v2 → · Published board — Porting projects to another language, including COBOL modernization.

  • Vibe Code Bench (Vals Index v2)7181.6%

    Vibe Code Bench (Vals Index v2) v2 → · Published board — End-to-end app-building tasks.

  • DeepSWE v1.1developer's claim66.6%† self-reported by the developer

    DeepSWE v1.1 v1.1 → · Vendor-reported by OpenAI — DeepSWE v1.1 score OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • Vibe Code Bench v1.14581.6%

    Vibe Code Bench v1.1 v1.1 → · Published board — Vibe Code Bench v1.1 results published by Vals AI.

  • FrontierCode 1.1 Extended (Cognition)developer's claim56.1%† self-reported by the developer

    FrontierCode 1.1 Extended (Cognition) v1.1 → · Codex — Whether a maintainer would merge the agent's pull request, on the full 150-task FrontierCode set crafted by open-source maintainers and graded with tests, rubrics and verifiers.

  • ProgramBench360.5%

    ProgramBench published 2026-09-23 → · Published board — ProgramBench results published by Vals AI.

  • Code Migration (standalone Vals board, 2026-09-22 snapshot)5042.6%

    Code Migration (standalone Vals board, 2026-09-22 snapshot) published 2026-09-22 → · Published board — Standalone Code Migration results published by Vals AI, distinct from its Vals Index v2 subset.

  • LMArena WebDev (Code Arena, overall)401,593 Elo

    LMArena WebDev (Code Arena, overall) published 2026-09-26 → · Codex harness — Community pairwise-vote ranking of models on front-end web development tasks, including agentic multi-step coding, published by LMArena as a Bradley-Terry rating.

Efficiency

  • FrontierCode 1.1 Main cost per rollout (Cognition)developer's claim$0.1† self-reported by the developer

    FrontierCode 1.1 Main cost per rollout (Cognition) v1.1 → · Codex — Cognition's published mean USD spend per rollout for each FrontierCode 1.1 Main model and reasoning effort.

  • Vals Index v2 cost per test (Vals AI)89$0.42

    Vals Index v2 cost per test (Vals AI) v2 → · Published board — Vals AI's published USD cost per test for each model on the Vals Index v2.

  • AutomationBench 1.0.6 cost per taskdeveloper's claim$0.037† self-reported by the developer

    AutomationBench 1.0.6 cost per task v1.0.6 → · Published board — USD cost per AutomationBench 1.0.6 task OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • Cost to run AA Intelligence Index (total) v4.3.276$122

    Cost to run AA Intelligence Index (total) v4.3.2 → · Published board — Cost to run AA Intelligence Index (total) results published by Artificial Analysis.

  • Cost per AA Intelligence Index task v4.3.2760.07

    Cost per AA Intelligence Index task v4.3.2 → · Published board — Cost per AA Intelligence Index task results published by Artificial Analysis.

  • FrontierCode 1.1 Extended cost per rollout (Cognition)developer's claim$0.083† self-reported by the developer

    FrontierCode 1.1 Extended cost per rollout (Cognition) v1.1 → · Codex — Cognition's published mean USD spend per rollout for each FrontierCode 1.1 Extended model and reasoning effort.

  • Agents' Last Exam V1 cost per taskdeveloper's claim$0.15† self-reported by the developer

    Agents' Last Exam V1 cost per task v1 → · Published board — USD cost per Agents' Last Exam V1 task OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • DeepSWE v1.1 cost per taskdeveloper's claim$0.22† self-reported by the developer

    DeepSWE v1.1 cost per task v1.1 → · Published board — USD cost per DeepSWE v1.1 task OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

  • OSWorld 2.0 offline (v2026.08.08 release) cost per taskdeveloper's claim$0.27† self-reported by the developer

    OSWorld 2.0 offline (v2026.08.08 release) cost per task v2026.08.08 → · Published board — USD cost per OSWorld 2.0 offline (v2026.08.08 release) task OpenAI reports for its own GPT-6 models in the 2026-09-22 GPT-6 Sol and Luna launch post.

Knowledge

Long-context

  • AA-LCR v1.19783.3%

    AA-LCR v1.1 v1.1 → · Published board — Tests reasoning across multiple long documents with corrected answer keys and grading.

Math

Reasoning

  • ARC-AGI-12286.7%

    ARC-AGI-1 v1 → · Published board — The first-generation ARC-AGI benchmark measuring passive fluid intelligence, with leaderboard score plotted against cost-per-task.

  • ARC-AGI-22259.3%

    ARC-AGI-2 v2 → · Published board — The second-generation ARC-AGI benchmark of passive fluid intelligence, with leaderboard score plotted against cost-per-task.

  • LiveBench 2026-06-252072.0

    LiveBench 2026-06-25 → · Published board — A periodically refreshed suite of objectively graded tasks with separate category scores and an overall category mean.

  • AA Intelligence Index v5.0.1 · snapshot-2026-11-044227.9

    AA Intelligence Index v5.0.1 · snapshot-2026-11-04 → · Published board — Artificial Analysis Intelligence Index v5.0.1. Models AA has not rescored on v5.0.1 are listed as estimates from their frozen v4.x value (isotonic fit over the models measured on both versions), never as measurements.

Science

Tool-use

Vision

  • MMMU Pro (AA)7175.5%

    MMMU Pro (AA) published 2026-09-10 → · Published board — Tests multimodal understanding using challenging ten-option questions.

  • Chartography 100-tasks5129.1%

    Chartography 100-tasks → · Published board — Can a model read the charts professionals rely on — Kaplan-Meier curves, candlesticks, contour maps, Sankey diagrams, Bode plots, wind roses — from the image alone, to an expert-set precision?

  • GDP.pdf 100-tasks4923.0%

    GDP.pdf 100-tasks → · Published board — Can a model answer real professional questions about the PDFs work runs on — manuals, wiring diagrams, filings, forms — so that every rubric criterion an expert wrote is met?

  • Furniture Assembly Benchmark (Epoch AI)2328.3%

    Furniture Assembly Benchmark (Epoch AI) published 2026-09-26 → · Published board — Spatial-reasoning test by Epoch AI: given a photo of a partly or fully assembled IKEA product and its official manual, the model must say whether the build is correct and, if not, name the step with the mistake.

Unusual results

1 threshold-crossing signal flagged · 21 eligible benchmark families

How flags are calculated

Heuristic screen, not statistical significance: benchmark families are correlated and source uncertainty is unknown. Peer evidence needs ≥ 20 independently measured matched configurations from ≥ 10 distinct model families. A flag needs a directed population z-score of magnitude ≥ 1.5 and a gap of ≥ 1.5 from the leave-one-benchmark-family-out mean z in the same direction, over ≥ 5 other benchmark families.

  • unusually strong CritPt (AA) published 2026-09-10
    Why
    Observed: 0.19429 fraction
    Peer mean 0.04412 · peer sd 0.08179 · n 537 · families 370
    Directed z 1.836 · baseline z 0.267 · gap 1.569 · profile n 20
    0.19429 fractionmeasuredobserved 2026-09-22artificialanalysis.ai ↗
    Evidence
    Axis: CritPt (AA) · published 2026-09-10 · Published board
    Exact value: 0.194285714285714 fraction
    Observed: 2026-09-22T22:47:42.324143+00:00 · publication date: not recorded
    Observation id: cr128:6f0a1242af2507ad5cbd8565
Missing coverage · 267 benchmark versions

No result does not mean a zero, or that the model was never tested. Collection failures and disputed versions retain their distinct status.

Variants / reasoning settings

Artificial Analysis snapshot 2026-11-04 · Intelligence Index v5.0 (est. = estimated from v4, not measured) · Coding Index v4.x frozen 2026-10-27 · Data: Artificial Analysis · Data: Epoch AI (CC BY)

VariantAA CodingAA Intelligence
GPT-6 Luna (Max)—27.9
GPT-6 Luna (Xhigh)—23.5
GPT-6 Luna (High)—21.4
GPT-6 Luna (Medium)—18.5
GPT-6 Luna (Low)—9.7
GPT-6 Luna (Non-reasoning)—9.4

SUBSCRIPTION PLAN

GitHub Copilot

Current usage-based billing · model token cost is converted to AI Credits at 1 credit = $0.01.

Input $0.100 / 1M
Cached input $0.010 / 1M
Cache write $0.125 / 1M
Output $0.500 / 1M
Status GA
Token offers by platform · 7 offers (Adjusted $/task)

Click any underlined price to see how it is estimated and where each input comes from. How we calculate adjusted cost.

7 offers within the active global filters, 2 more outside them; “price n/a” means the provider serves the model but publishes no direct token price we could collect — no other channel's price is shown in its place.

OpenRouter (5)

Prices via OpenRouter — what OpenRouter charges for each provider's endpoint, not the provider's own direct price. Source: openrouter.ai/openai/gpt-6-luna/providers, collected 2026-11-05.

Azureconditional
Azureconditional
Amazon Bedrockconditional
Azureconditional
Amazon Bedrockconditional

T-Systems LLM Hub (2)

Direct prices from T-Systems LLM Hub's own price list. Source: docs.llmhub.t-systems.net/models/llms/, collected 2026-11-05.

T-Systems LLM Hub
T-Systems LLM Hub
GPT-6 Luna (Max) — benchmarks & cost | Benchmark Heaven