Model comparison
GPT-5.6 Luna vs GPT-6 Astra
Two OpenAI tiers compared on the figures that decide which one a workload actually needs.
Catalog record checked September 5, 2026Individual provider fields may changeEvidence confidence: High — see receipts below
OpenAI
GPT-5.6 Luna
Budget
OpenAI
GPT-6 Astra
Frontier
| Specification | GPT-5.6 Luna | GPT-6 Astra |
|---|---|---|
| Provider | OpenAI | OpenAI |
| Tier | Budget | Frontier |
| Context window | 1.05M | 1.05M |
| Max output | 128K | 128K |
| Input / 1M tokens | Winner: $0.20 | $10 |
| Output / 1M tokens | Winner: $1.20 | $50 |
| Weights | Closed | Closed |
| Parameters | cheapest GPT-5.6 tier | computer-use flagship |
| Reasoning levels | none, low, medium, high | low, medium, high, xhigh, max |
| Modalities | text, image | text, image |
| API model id | gpt-5.6-luna | gpt-6-astra |
| Released | July 9, 2026 | September 3, 2026 |
| Artificial Analysis Intelligence Index [high] (2026-08-14) | 47 | Not verifiedUnverified |
| Artificial Analysis Intelligence Index [max] (2026-09-03) | Not verifiedUnverified | 61.2 |
| SWE-bench Verified (2026-09-01) | 93 | Not verifiedUnverified |
| Terminal-Bench 2.1 (2026-07-09) | 84.7 | Winner: 89.9 |
| LiveBench (2026-09-05) | 73.6 | Not verifiedUnverified |
| DeepSWE 1.1 (2026-09-03) | Not verifiedUnverified | 74.1 |
| OSWorld 2.0 (offline) (2026-09-03) | Not verifiedUnverified | 72.6 |
| GPQA Diamond (2026-09-03) | Not verifiedUnverified | 96 |
| Humanity's Last Exam (2026-09-03) | Not verifiedUnverified | 57.2 |
Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Bold indicates the better figure where one is objectively better. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.
Benchmark receipts
- 2026-09-05: Artificial Analysis Terminal-Bench v2.1 (independent, Terminus 2 harness — conflicts with tbench.ai 87.4; harness difference preserved)Agentic terminal work: multi-step tasks executed in a sandboxed shell environment. Comparability: comparable with caveat — Scores across Terminal-Bench major versions (v1 vs v2) are NOT comparable; harness configuration (container, time limits) also shifts results.
- 2026-09-05: LiveBench official leaderboard (benchmark-owned), max effortContamination-resistant general capability across reasoning, coding, math, data analysis, and language, with monthly question refreshes. Comparability: comparable with caveat — Rolling question set: observations months apart measure different question mixes. Record the measurement date and compare within ~1 month windows.
- 2026-09-03: Terminal-Bench official leaderboard (independent, native harness, high effort; captured 2026-09-03 — the live board has since moved to Terminal-Bench 4.0)Agentic terminal work: multi-step tasks executed in a sandboxed shell environment. Comparability: comparable with caveat — Scores across Terminal-Bench major versions (v1 vs v2) are NOT comparable; harness configuration (container, time limits) also shifts results.
- 2026-09-03: OpenAI GPT-6 Astra launch chart (vendor, with tools; transcribed by Vellum)Agentic software engineering: pass@1 on real repo tasks with reported cost per completed task. Comparability: directly comparable
- 2026-09-03: Artificial AnalysisComposite index blending reasoning, knowledge, and coding evals into one 0–100 score. Comparability: directly comparable — AA occasionally rebaselines the index scale between snapshots — a score captured on one date is only comparable to same-snapshot scores (check measuredAt).
- 2026-09-01: vals.ai SWE-bench Verified leaderboard (independent, mini-swe-agent bash-only) (independent; OpenAI did not report SWE-bench Verified for the GPT-5.6 family)Real GitHub issue resolution: does the model's patch pass the hidden tests. Comparability: comparable with caveat — Post-audit vendor claims and pre-audit scores sit on different task trust levels; scaffolding (agent harness, compute budget) also dominates results. Never aggregate across scaffolds.
- 2026-07-09: OpenAI GPT-5.6 announcement (vendor; chart transcribed by Vellum)Agentic terminal work: multi-step tasks executed in a sandboxed shell environment. Comparability: comparable with caveat — Scores across Terminal-Bench major versions (v1 vs v2) are NOT comparable; harness configuration (container, time limits) also shifts results.
Pricing tiers: GPT-5.6 Luna: Launch list was $1/$6; cut to $0.20/$1.20 on 2026-07-30. Cached input $0.02/MTok; batch $0.10/$0.60. · GPT-6 Astra: Standard $10/$50 per MTok (≤272K context); >272K input billed $20/$75 (Fast mode output $100). Cached input $1/MTok; Fast mode is 2x price for ~2.5x speed. Knowledge cutoff Apr 30 2026. Rolling out over several days from 2026-09-03; Enterprise tenants get it disabled by default.
GPT-5.6 Luna
GPT-5.6 Luna is OpenAI's cheapest tier — roughly 1/25 of Opus 5-class input cost, suitable where quality requirements are modest.
Best for
- Classification
- High-volume chat
- Latency-sensitive pipelines
Watch out
Smallest GPT-5.6 tier; verify quality holds before routing frontier work to it.
GPT-6 Astra
GPT-6 Astra is OpenAI's computer-use flagship — the first model it gates at the Preparedness Framework's 'Critical' cybersecurity threshold, trained on >100,000 GPUs at Stargate.
Best for
- Computer-use agents
- Agentic coding
- Hard reasoning
Watch out
Gated at the 'Critical' cyber threshold and disabled by default for Enterprise; long-context work bills $20/$75 above 272K input tokens.
When the cheaper one wins
GPT-5.6 Luna is cheaper on output at $1.20 per million tokens against $50 for GPT-6 Astra — about 42×. Use the cheaper tier for classification, extraction, summarisation, and any task where the expensive model’s extra score does not change the accepted output. The expensive one only pays if your hardest task actually fails on the cheap tier. These are standard-tier API rates, excluding batch and cache discounts.
Run the model pickerEvidence confidence: High
How strong and complete the evidence behind this comparison is — not a prediction of which model is better.
- Pricing verified on both sides — Input and output rates are verified for both models.
- 5/5 core specs verified on both sides — All core specifications verified for both models.
- 2 shared named benchmarks with differing scores — Measured on: Artificial Analysis Intelligence Index, Terminal-Bench 2.1.
- Verified within the last 90 days — Newest catalog check was 6 days ago.
- Both models carry source citations — Each side has at least two catalog sources on record.
Source receipts
Each catalog figure was checked against the provider or an independent second source on the date shown.
- GPT-5.6 Luna: OpenAI — GPT-5.6 (accessed 2026-08-29)
- GPT-5.6 Luna: OpenAI API pricing (gpt-5.6-luna $0.20/$1.20) (accessed 2026-08-29)
- GPT-6 Astra: OpenAI — Introducing GPT-6 Astra (accessed 2026-09-03)
- GPT-6 Astra: OpenAI API pricing (gpt-6-astra $10/$50) (accessed 2026-09-03)
- GPT-6 Astra: CNBC — OpenAI launches GPT-6 Astra (accessed 2026-09-03)
- GPT-6 Astra: OpenAI model docs — gpt-6-astra (128K max output, effort levels) (accessed 2026-09-05)
Related comparisons
- Claude Opus 5 vs GPT-5.6 Luna
- DeepSeek V4 Flash vs GPT-5.6 Luna
- Gemini 3.6 Flash vs GPT-5.6 Luna
- Gemini 3.7 Flash vs GPT-5.6 Luna
- GPT-5.6 Luna vs Grok 4.6
- Claude Fable 5.1 vs GPT-6 Astra
Diving deeper on one model? GPT-5.6 Luna · GPT-6 Astra
Common questions
GPT-5.6 Luna vs GPT-6 Astra
Answered from the verified figures on this page rather than general guidance.
Is GPT-5.6 Luna or GPT-6 Astra cheaper for input?
GPT-5.6 Luna is cheaper at $0.20 per million input tokens, against $10 for GPT-6 Astra — roughly 50× the price. Output tokens usually dominate a real bill, so weigh the output rate more heavily than the input rate. These are base rates; GPT-5.6 Luna has tiered pricing: Launch list was $1/$6; cut to $0.20/$1.20 on 2026-07-30. Cached input $0.02/MTok; batch $0.10/$0.60. GPT-6 Astra has tiered pricing: Standard $10/$50 per MTok (≤272K context); >272K input billed $20/$75 (Fast mode output $100). Cached input $1/MTok; Fast mode is 2x price for ~2.5x speed. Knowledge cutoff Apr 30 2026. Rolling out over several days from 2026-09-03; Enterprise tenants get it disabled by default.
Is GPT-5.6 Luna or GPT-6 Astra cheaper for output?
GPT-5.6 Luna is cheaper at $1.20 per million output tokens, against $50 for GPT-6 Astra — roughly 42× the price. Output tokens usually dominate a real bill, so weigh the output rate more heavily than the input rate. These are base rates; GPT-5.6 Luna has tiered pricing: Launch list was $1/$6; cut to $0.20/$1.20 on 2026-07-30. Cached input $0.02/MTok; batch $0.10/$0.60. GPT-6 Astra has tiered pricing: Standard $10/$50 per MTok (≤272K context); >272K input billed $20/$75 (Fast mode output $100). Cached input $1/MTok; Fast mode is 2x price for ~2.5x speed. Knowledge cutoff Apr 30 2026. Rolling out over several days from 2026-09-03; Enterprise tenants get it disabled by default.
Which has the larger context window, GPT-5.6 Luna or GPT-6 Astra?
Both accept about 1.05M tokens of context, so document length will not decide between them.
Do GPT-5.6 Luna and GPT-6 Astra support the same reasoning levels?
GPT-5.6 Luna exposes none, low, medium, high, while GPT-6 Astra exposes low, medium, high, xhigh, max.
Should I use GPT-5.6 Luna or GPT-6 Astra?
GPT-5.6 Luna is the budget tier and GPT-6 Astra the frontier tier. The useful question is whether your hardest task actually fails on the cheaper one — most production volume such as classification, extraction and summarisation does not.
Next step
Choosing between them
The verified figures that separate this pair, computed from the catalog rather than restated boilerplate.
- Input price: GPT-5.6 Luna costs $0.20 per 1M tokens versus $10 for GPT-6 Astra — a 50x difference at the headline tier.
- Measured capability: GPT-6 Astra leads Artificial Analysis Intelligence Index 61.2 to 47 (measured 2026-08-14).
- Positioning: GPT-5.6 Luna sits in the budget tier, GPT-6 Astra in the frontier tier — most production volume (classification, extraction, summarisation) does not need the pricier tier.
Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Windsurf, or Lovable? Compare agentic harnesses · Latest releases.