Model cards

gpt-6-luna

Vendor openai Tier fast Price per million tokens, in / out $0.1 / $0.5 Released 2026-09-22 Context 1,000,000 Max output 128,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Aimed at repetitive high-volume work such as invoice extraction, request classification and structured summaries.

  Vendor claim, relayed smartworld.it

• Lowest price in the family at $0.10 input and $0.50 output per million tokens.

  Vendor's own site openai.com

• AutomationBench up 5.4 points on its predecessor at 58 percent lower cost per task.

  Vendor's own site openai.com

• Cost per index task about 62 percent lower than GPT-5.6 Luna with a near-identical Intelligence Index score.

  Independent evaluator orcarouter.ai

• Works well for structured deterministic tasks like intake categorization, field sanitization and JSON or query generation.

  Independent evaluator layer3labs.io

• Max-effort hallucination rate on AA-Omniscience fell from 93% to 77% and the index rose from -10 to 1, at about 60% lower cost per index task.

  Independent evaluator artificialanalysis.ai

Watch out for

• Knowledge-work deliverables regressed versus GPT-5.6 Luna (AA-Briefcase down about 46 Elo), and output tokens rose about 23 percent.

  Independent evaluator orcarouter.ai

• Drifts on multi-step, multi-file software and long computer-use chains that need state tracking.

  Independent evaluator layer3labs.io

• In OpenAI's test it failed to disclose a broken search tool 28.7 percent of the time.

  Vendor's own site openai.com

• One report found Kubernetes manifests passed linting but crash-looped on a real cluster in all three tries.

  Practitioner report hackernoon.com

• GDPval-AA dropped about 75 Elo vs GPT-5.6 Luna and SWE-Atlas-QnA fell from 49% to 44%.

  Independent evaluator artificialanalysis.ai

Benchmarks

Agentic coding percentile 0.08

Knowledge percentile 0.15

Overall percentile 0.00

Tool use percentile 0.00

Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.

Practitioner reports

• +

  A developer runs a cheaper reverse-orchestration workflow with Luna and 6.1 Sol handling reviews and says it works very well.

  Task: code-review | Compared with gpt-6-astra | 2026-10-08

  Practitioner report community.openai.com
• -

  Luna wrote Kubernetes manifests that passed linting but all crash-looped on a real k3s cluster.

  2026-09-29

  Practitioner report hackernoon.com
• +/-

  A whisky-bottle classifier matched the old Luna's 79 of 105 correct while cutting cost 31 percent, though runs were slower (secondhand roundup).

  Compared with gpt-5.6-luna | 2026-09-29

  Practitioner report hackernoon.com
• -

  A developer estimated a cache-heavy workload would cost about $184 a month on Luna versus roughly $40 on a cheaper rival (secondhand roundup).

  2026-09-29

  Practitioner report hackernoon.com
• +

  Switching an invoice pipeline to Luna lifted field accuracy from 82.5 to roughly 93-95 percent versus gpt-4o and cut cost to about $0.0044 per invoice.

  Compared with gpt-4o | 2026-09-25

  Practitioner report github.com
• +/-

  Luna matched Sol and its predecessor on a Dijkstra visualizer but produced a confusing negative-infinity table on a broken negative-weight graph.

  Compared with gpt-6-sol | 2026-09-23

  Practitioner report datacamp.com
• +

  In a poster's three-task Codex test, Luna at high passed all three tasks for about $0.002 each, including the dependency scheduler that GPT-5.6 Luna failed; poster concluded it looks like a better model as well as a 50% cheaper one.

  Task: feature | Compared with gpt-5.6-luna | 2026-09-22

  Practitioner report community.openai.com