OpenAI
GPT-6 Luna
The cheapest GPT-6 at $0.10/$0.50 per million tokens — half GPT-5.6 Luna's promotional input rate and 42% of its output rate. On independent data it ties GPT-5.6 Luna on Humanity's Last Exam (38.5% vs 39.5%), LiveBench (72.0 vs 73.6) and ARC-AGI-2 (59.31% vs 59.6%), but trails it on Agents' Last Exam, 25.0 vs 30.3, a real 5.3-point gap.
GPT-6 Luna benchmarks and pricing, every number sourced: 9 tracked GPT-6 Luna benchmark scores (7 independently run, 2 still resting on a vendor’s own claim), priced at $0.10 per million input tokens and $0.50 per million output.
- Released
- 2026-09-22
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- 2026-05-18
- Verified
- sources checked 2026-09-22–2026-10-25
- Succeeded
- GPT-5.6 Luna (2026-07-09)
- Parameters
- Not disclosed
- Architecture
- Not disclosed
GPT-6 Luna’s verified record
GPT-6 Luna’s most-compared rival is GPT-6 Astra: 4 trails, 1 tie, and 3 not callable across their 8 shared comparisons. GPT-6 Luna is priced at $0.10/$0.50 per 1M tokens (in/out) vs GPT-6 Astra’s $10.00/$50.00.
Against the 227 head-to-head comparisons GPT-6 Luna shares with other tracked models:
A gap counts for GPT-6 Luna only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — GPT-6 Luna trails on 64 of them.
HLE · no tools Reasoning
±2 is noiseLiveBench Composite score across 7 domains
±2.7 is noiseTerminal-Bench 4.0 · max Terminal ops
±12.4 is noiseARC-AGI-2 · max Compositional visual reasoning
±9.2 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseAgents' Last Exam Professional work
±3.2 is noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseNo verdict for GPT-6 Luna anywhere on DeepSWE, OSWorld 2.0 · v2026 08 08 offline ( nothing independently confirmed on both sides).
- None of GPT-6 Luna’s coding comparisons are independently confirmed on both sides yet.
What changed from GPT-5.6 Luna to GPT-6 Luna
The 7 benchmarks both models have been scored on, using the same variant each time. A raw GPT-6 Luna gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.
| Benchmark | GPT-5.6 Luna | GPT-6 Luna | Change | Verdict |
|---|---|---|---|---|
| HLE | 39.5 | 38.5 | -1.0 | Tie |
| Terminal-Bench 2.1 | 80.9 | 73.03 | -7.9 | Setup-dependent |
| DeepSWE | 67 | 66.6 | -0.4 | Unverified |
| Agents' Last Exam | 30.3 | 25 | -5.3 | Setup-dependent |
| ARC-AGI-2 | 59.6 | 59.31 | -0.3 | Tie |
| LiveBench | 73.6 | 72 | -1.6 | Tie |
| Terminal-Bench 4.0 | 11.62 | 13.64 | +2.0 | Tie |
GPT-6 Luna API pricing
$0.10 in / $0.50 out per 1M tokens — official pricing source
What GPT-6 Luna costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.015 |
| A codebase review | 1,000K / 100K | $0.150 |
| A day of agent work | 10,000K / 1,000K | $1.50 |
Computed from GPT-6 Luna’s list rates above — cache discounts and batch tiers are not applied.
GPT-6 Luna is one of 8 OpenAI models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| GPT-6 Cyber | $20.00 | $100.00 |
| GPT-6.1 Terra | $1.00 | $5.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-6 Sol(previous version) | $2.00 | $10.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-5.6 Luna(previous version) | $0.20 | $1.20 |
| GPT-5.6 Sol(previous version) | $4.00 | $20.00 |
GPT-6 Luna benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| ARC-AGI-2(max)[2] Compositional visual reasoning · ±9.2 is noise | |
| LiveBench[3] Composite score across 7 domains · ±2.7 is noise | |
| DeepSWE[4] Long-horizon coding · ±9.5 is noise | |
| Terminal-Bench 2.1[5] Terminal ops · ±10.6 is noise | |
| Agents' Last Exam[6] Professional work · ±3.2 is noise | |
| Terminal-Bench 4.0(max)[7] Terminal ops · ±12.4 is noise | |
| AnalystAgent[8] Spreadsheet & document analysis · ±11.2 is noise | |
| OSWorld 2.0(v2026 08_08_offline)[9] Long-horizon computer use · ±9.7 is noise |
Who ran these numbers: 7 of 9 independent — artificialanalysis.ai (2), vals.ai (2), arcprize.org (1), livebench.ai (1), snorkel.ai (1); vendor self-reported (2).
- HLE: AA's own run of GPT-6 Luna at max effort, text-only 2,158-question subset.
- ARC-AGI-2: ARC Prize's own leaderboard row at the max effort tier ($0.062/task); v2.json dataset generated 2026-09-22.
- LiveBench: Board row 'GPT-6 Luna Max Effort' on the 2026-06-25 LiveBench release; only tier populated for this model.
- DeepSWE: OpenAI's own launch-post figure (max effort). No independent DeepSWE run for this model has appeared on deepswe.datacurve.ai or via Artificial Analysis's own agent-harness runs yet — unlike GPT-6 Sol, which AA has run through its Codex agent.
- Terminal-Bench 2.1: vals.ai's own Terminus 2 run, board row 'GPT-6 Luna' (73.03 ± 1.72, pass@1, reasoning effort max), from its archived Terminal-Bench 2.1 table. One provider refusal on 267 tasks (0.38%) is scored as a failure.
- Agents' Last Exam: Overall pass rate 25.0 (Codex, Max, the board's best effort for this model; score 50.9; est. cost $21). Per-effort results: Max 25.0, XHigh 24.3, High 21.1, Medium 20.4, Low 16.4. OpenAI's launch chart self-runs ALE at 50.9 max (V1 designation) — far above the board's 25 stored here; not stored as a row.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 13.636 at effort max. Artificial Analysis same-conditions run (Max) reads 12.63 against this 13.64. Official board (grant-funded, Codex, max effort, 5-trial pass@1): 16.36 ± 2.72 CI — 2.7 points above this run.
- AnalystAgent: AA's own run, board row 'GPT-6 Luna (Max)': analystAgent 0.4375 = 43.75 pass^5. pass@1 is 65.75. Read from the GPT-6 Luna (Max) page on 2026-10-15.
- OSWorld 2.0: Launch chart 'OSWorld 2.0, offline set' at max effort; same partial-reward / v2026.08.08-offline footnote as GPT-6 Sol's row.
Notes on the record
Pricing is tiered by prompt length, confirmed on developers.openai.com/api/docs/models/gpt-6-luna: $0.10 input / $0.50 output per million tokens under 272,000 prompt tokens (cached input $0.01), rising to $0.20/$0.75 (cached $0.02) — 2x input and cache, 1.5x output, billed on the full request — once a prompt crosses that threshold, per OpenAI's own wording. OpenAI's comparison table reads "$0.20→$0.10 input, $1.20→$0.50 output, 50% cheaper" against GPT-5.6 Luna, but calls that $0.20/$1.20 rate "promotional pricing" — GPT-6 Luna's own price is presented as standard, not a further discount on a discount.
Context window 1,050,000 tokens, max output 128,000. Input: text and images. Output: text only. Knowledge cutoff May 18, 2026. License not stated for GPT-6 Luna; inferred proprietary, API-only.
No OpenAI page calls this a replacement for GPT-5.6 Luna. Beyond the absence of "supersede"/"replace" language, GPT-5.6 Luna is explicitly still in active use elsewhere on OpenAI's own site — its API deprecations page lists GPT-5.6 Luna as the recommended replacement for two older, separately-deprecating models.
OpenAI self-reports 66.6% on DeepSWE v1.1 (max effort); no independent DeepSWE run for this specific model has appeared yet on deepswe.datacurve.ai or Artificial Analysis's own agent-harness runs (unlike GPT-6 Sol, which AA has run through its Codex-agent harness).
Two independent rows were added on 2026-10-01. Snorkel's Agents' Last Exam board lists GPT-6 Luna at 25.0 overall pass rate (Codex; the board defaults to Max), 5.3 points under GPT-5.6 Luna's 30.3 (at matched Max, 25.0 vs 28.3; at XHigh, 24.3 vs 30.3), past the 3.2-point band either way: the one real gap between the two. vals.ai's Terminal-Bench 2.1 run reads 73.03 at max effort, with one provider refusal scored as a failure; GPT-5.6 Luna's row there is Artificial Analysis's, so that pair is setup-dependent, and on vals.ai's own board the two read 79.03 and 73.03, a tie.
Compare with
FAQ
Has GPT-6 Luna been independently benchmarked?
Partly. Seven of the fourteen benchmarks this site tracks carry an independently-run score: Humanity's Last Exam (38.5%, Artificial Analysis, max effort), ARC-AGI-2 (59.31%, ARC Prize, max tier) and LiveBench (72.0 overall, max effort) from the day after launch, plus vals.ai's Terminal-Bench 2.1 run (73.03, max effort) and Snorkel's Agents' Last Exam (25.0 overall pass rate, Codex), both recorded 2026-10-01. Added 2026-10-08: vals.ai's Terminal-Bench 4.0 run (13.64, max effort). Added 2026-10-15: AA-AnalystAgent 43.75 pass^5 at max effort (Artificial Analysis). GPQA Diamond, SWE-bench Verified, LiveCodeBench, DeepSWE, Toolathlon-Verified and HMMT Feb 2026 had none when checked; the DeepSWE figure here is OpenAI's own.
Is GPT-6 Luna better than GPT-5.6 Luna?
No, and on one benchmark it is worse. Humanity's Last Exam (38.5% versus 39.5%), LiveBench (72.0 versus 73.6) and ARC-AGI-2 (59.31% versus 59.6%), each from the same evaluator at max effort, are ties inside their noise bands. Snorkel's Agents' Last Exam, where the board shows each model at its best effort, is a real gap the other way: 25.0 versus 30.3, past the 3.2-point band. OpenAI's own improvement claim for Luna, 5.4 points on AutomationBench, is on a benchmark this site doesn't track. What clearly changed is price: $0.10/$0.50 versus $0.20/$1.20.
Does GPT-6 Luna's price change with longer prompts?
Yes. The $0.10 input / $0.50 output per-million-token rate applies to prompts up to 272,000 tokens. Once a prompt exceeds that threshold, OpenAI bills the whole request at 2x the input and cached-input rate and 1.5x the output rate — $0.20 input / $0.02 cached / $0.75 output — per developers.openai.com, confirmed 2026-09-23.
Does GPT-6 Luna replace GPT-5.6 Luna?
No. OpenAI's own API deprecations page lists GPT-5.6 Luna as the recommended replacement for two other, older models that are themselves being retired — meaning OpenAI is actively steering traffic toward GPT-5.6 Luna elsewhere on the same site, not away from it. Neither model's page uses "supersede" or "replace" to describe the relationship between them.
How does GPT-6 Luna compare to GPT-6 Sol?
Both launched the same day at the same tiered-pricing structure, but Luna is roughly a twentieth of Sol's price ($0.10/$0.50 versus $2/$10) and trails on every independent benchmark this page tracks for both: real gaps on Humanity's Last Exam (38.5% versus 47.9%), LiveBench (72.0 versus 79.3 overall) and Agents' Last Exam (25.0 versus 32.2), and a 10.1-point Terminal-Bench 2.1 deficit on vals.ai's runs (73.03 versus 83.15) that sits just inside the 10.6 band. OpenAI positions Luna as the cheaper, smaller sibling in the same GPT-6 family as Sol and Astra.
What is GPT-6 Luna's knowledge cutoff?
May 18, 2026, with a 1,050,000-token context window. Within the GPT-6 family tracked here, its verified strength so far is terminal work: Terminal-Bench 2.1 at 73.03 (2026-10-01) is the best current-benchmark figure on this page.
Further reading
- Models with 10M token context windows 2026 — GPT-6 Luna is one of the 47 models it compares.
Benchmark guides
- Humanity's Last Exam — what GPT-6 Luna’s reasoning score on it does and doesn’t prove.
- ARC-AGI-2 — what GPT-6 Luna’s reasoning score on it does and doesn’t prove.