GPT-5.6 Luna
Language Available Full comparison ↗ openai/gpt-5.6-luna · by OpenAI
· decoder-only-transformer
Pricing — 1 offering(s)
About this offering: OpenAI GPT-5.6 Luna -- the cost-optimized tier of the GPT-5.6 family ("cost-sensitive, high-volume workloads" per OpenAI's model docs), launched 2026-07-10 alongside Sol and Terra.
Input tokens
- $1.00 / 1M tokens (input) Historical 2026-07-10 → 2026-09-19
- $0.20 / 1M tokens (input)
Current
2026-09-20 → present
OpenAI cut Luna's input price from $1.00 to $0.20 per 1M tokens on 2026-09-20. This is the short-context rate…
OpenAI cut Luna's input price from $1.00 to $0.20 per 1M tokens on 2026-09-20. This is the short-context rate (up to 272K input tokens); long-context input is $0.40 per 1M (OpenAI pricing page, checked 2026-10-10).
Output tokens
- $6.00 / 1M tokens (output) Historical 2026-07-10 → 2026-09-19
- $1.20 / 1M tokens (output)
Current
2026-09-20 → present
OpenAI cut Luna's output price from $6.00 to $1.20 per 1M tokens on 2026-09-20. This is the short-context…
OpenAI cut Luna's output price from $6.00 to $1.20 per 1M tokens on 2026-09-20. This is the short-context rate; above 272K input tokens, output is $1.80 per 1M.
Batch
Batch input tokens
- $0.10 / 1M tokens (input) Current 2026-11-03 → present
Batch output tokens
- $0.60 / 1M tokens (output) Current 2026-11-03 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how GPT-5.6 Luna fits into a cost-aware routing setup
See how →Capability profile
How GPT-5.6 Luna rates across core capability dimensions, with the task-level evidence behind each rating.
OpenAI's official model spec page (developers.openai.com/api/docs/models/gpt-5.6-luna) lists 'Reasoning: High' — the lowest of a distinct three-way Highest/Higher/High ranking across the family (Sol=Highest, Terra=Higher, Luna=High), which does distinguish this tier unlike the identical 'Fast' speed label shared by all three. Rated 'weak' relative to Sol's 'strong' (Highest) and Terra's 'moderate' (Higher) — a within-family comparison, not an absolute judgment: OpenAI's own label for this tier is still 'High', not low.
SWE-bench Verified (500 human-verified GitHub issues; the patch must pass the repo's own tests; no file-location hints): resolves 93%, so it fails roughly 1 issue in 14. Independent (Vals AI leaderboard; OpenAI's own launch page publishes no SWE-bench figure for this tier).
OpenAI API docs (developers.openai.com/api/docs/models): supports Functions, Web search, File search, Computer use — same tool set as Sol and Terra.
1.05M-token context window (developers.openai.com/api/docs/models/gpt-5.6-luna) — same family-wide window as Sol and Terra.
OpenAI's own GPT-5.6 launch page (openai.com/index/gpt-5-6/, 'Availability and pricing' section, live-browser-verified 2026-07-20 — flat HTTP fetch of this page 403s, same as the coding-dimension quote above): 'Luna, our fastest and most affordable model' — an explicit, distinct per-tier speed claim, unlike the identical family-wide 'Fast' label the per-model spec pages show for all three tiers. Rated 'strong' as the fastest of the three.
OpenAI API docs: explicitly described as 'optimized for cost-sensitive workloads' — cheapest tier in the family at $1/$6 per 1M tokens.
Ratings are estimated — limited independent data is available for this model.
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| swe-bench-verified ↗ | 93.0% 2026-07 | — | 📊 leaderboard ↗ |
| swe-bench-pro | 62.7% 2026-07 | — | 🏢 vendor ↗ |
| terminal-bench-2-1 | 84.7% 2026-07 | — | 🏢 vendor ↗ |
| terminal-bench-4-0 | 11.6% 2026-09 | mini-swe-agent | 📊 leaderboard ↗ |
Operator guidance
The cost-optimized tier of the GPT-5.6 family (launched 2026-07-10 alongside Sol and Terra). Route here for throughput and cost; step up to Terra or Sol when a task needs more reasoning or coding depth than Luna reliably provides.
Use cases
- Cost-sensitive, high-volume workloads — OpenAI's own positioning for this tier
Limitations
- instruction-following and multilingual ratings omitted — re-checked 2026-07-20 (live browser fetch of openai.com/index/gpt-5-6/ in full, the OpenAI Help Center's 'GPT-5.6 in ChatGPT' article, and the GPT-5.6 system card at deploymentsafety.openai.com/gpt-5-6) and still no distinct per-tier claim exists on any OpenAI-owned page for either dimension; rather than infer a rating, this doc leaves them out
- speed rating added 2026-07-20 (previously omitted) — a fuller re-read of the GA launch page's 'Availability and pricing' section (not just the per-tier spec pages the original pass checked) found 'Luna, our fastest and most affordable model,' a genuine distinguishing claim the original pass had missed; superseding the earlier note about the dropped preview-post adjective, which is no longer the operative source
- Capability ratings are qualitative, not a single cited benchmark run