GPT-6 Luna
Released Sep 22, 2026 · knowledge cutoff 2026-05
- Status
- Unknown
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1.1M
- Max output
- 128K
- Speed (medium effort)
- 143 tok/s · 5.28s to first answer token
- Price ($/MTok in / out)
- $0.10 / $0.50
- Cost per task (medium effort)
- $0.020
- Open weights
- No
Benchmarks2 of 10 reported
Humanity's Last Exam
ARC-AGI-2
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
LMArena Elo
GDPval-AA v2
Strengths & weaknesses
Strengths
- Competent at both SQL and front-end HTML/JavaScript inside an agent harness — Simon Willison swapped it into his Datasette agent demo on launch day
- Reasoning effort runs from none to max, so one model covers routing and classification at the bottom and heavier extraction at the top
- Fewer misleading claims about coding work than 5.6 Luna made, per OpenAI's own alignment stress tests
- Streams a little faster than 5.6 Luna at the same effort, and the low setting is the quickest thing in the GPT-6 line
Weaknesses
- Thinks longer to get there than 5.6 Luna did — Artificial Analysis measured about a quarter more output tokens per index run, which eats into the per-token price cut
- Falls apart on long terminal-driven agent tasks where Sol holds up — the gap between the two is far wider there than on plain coding
- Early users read it as a price cut rather than a capability step over 5.6 Luna, and independent indexes put the two level
Budget tier of the GPT-6 family. Context 1,050,000 tokens (922K max input), 128K output, cutoff 2026-05-18, and $0.10/$0.50 per MTok ($0.01 cached input, $0.125 cache writes) are from OpenAI's model docs; prompts over 272K input tokens bill at 2x input/cache and 1.5x output, a fast tier costs 2x, and OpenAI calls the rates permanent. predecessorId stays null: OpenAI's announcement says only that Luna "also improves substantially" over its GPT-5.6 counterpart, lineage wording rather than a replacement claim, and GPT-5.6 Luna stays on sale. OpenAI's launch table uses only untracked suites (AutomationBench 1.0.6 20.7%, Agents' Last Exam 50.9%, FrontierCode 1.1 42.4%, DeepSWE v1.1 66.6%, OSWorld 2.0 offline 52.7%, factual error rate 7.6%, all at max) and the model shipped without a system card or Preparedness rating; OpenAI's stress tests put its deception rate at 2.8% (5.6 Luna 9.5%) and failure-to-disclose at 30.2% (78.3%). ARC-AGI-2 59.3% is ARC Prize-verified at max effort (xhigh 41.9%, high 31.4%, medium 18.1%, low 4.6%, none 0.0%), published 2026-09-22 at arcprize.org/results/openai-gpt-6-luna; cost per task not stated there. HLE 39 is Artificial Analysis' no-tools run at max effort, shown to integer precision on AA's comparison page (level with 5.6 Luna). AA's Terminal-Bench run is 4.0 (13% max, not the tracked 2.1), and its GDPval-AA v2.1 ratings — 1367 max / 1218 medium, down from 5.6 Luna's 1443 — are on a different scale from the v2 figure this key records, so gdpvalAA stays null. AA Intelligence Index 37 at max ($0.07 per task) and 29 at medium, the effort the costPerTask and speed figures are taken at; AA measured 51K output tokens per index run against 5.6 Luna's 41K and a hallucination rate down from 93% to 77%. Not on the arena.ai text board (last update 2026-09-13) and absent from Scale's SWE-bench Pro board. Generally available in the API, Bedrock, GitHub Copilot and ChatGPT Work/Codex, with Free and Go users getting it in the desktop app.
For developers
API model strings
- openai
gpt-6-luna
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor