none reviewed yet · 2 benchmark observations · latest capture Sep 8, 2026
Inspect the evidenceOpenAI · MODEL PROFILE
GPT-5.6 Luna
Cost-sensitive teams running lots of code edits and CI-style fixes
Saved editorial fit for your current task lens; not a benchmark result
How scoring worksVerified rates (Sep 8, 2026) · 100 runs / month at your token settings
Adjust workloadAt a glance
9 facts- Provider
- OpenAI
- Type
- API model
- Access
- OpenAI API; also used as the judge model in Z.ai's GLM-5.3 evaluations
- Context
- 1,050,000 tokens
- Maximum output
- 128,000 tokens
- License / weights
- Proprietary
- Input modalities
- file, image, text
- Output modalities
- text
- Pricing notes
- $0.20 in / $1.20 out per 1M; cached input $0.02
The practical trade-offs
Documented strengths
- SWE-bench Pro 62.7% at $0.20/$1.20 is an extreme price/quality ratio
- 1.05M context at budget pricing
- 127.2 tok/s throughput
- Cached input $0.02 makes repeat-context agents very cheap
Limitations to check
- Arena frontend Elo 1516 is near the bottom of this cohort
- TTFT 124.70 s at max effort
- AA Intelligence Index 43 is well below the frontier
- No visual/design benchmark evidence
Follow the source
5 source linksEvidence & conditions
5Source links
NoneReviewed records
2Benchmark runs
Sep 8, 2026Latest capture
- Evidence source 1
developers.openai.comDate not recorded - Evidence source 2
openai.comDate not recorded - Evidence source 3
arena.aiDate not recorded - Evidence source 4
benchlm.aiDate not recorded - Evidence source 5
artificialanalysis.aiDate not recorded
Benchmark observations
SWE-Bench Pro Inspect conditions
- name
- SWE-Bench Pro
- value
- 62.7
- unit
- %
- source Url
- https://openai.com/index/gpt-5-6/
- checked At
- 2026-09-08T19:30:56Z
- attribution
- provider-reported
- configuration
- OpenAI launch table; single-model column, not Ultra. Exact effort and harness not stated in the table.
DeepSWE v1.1 Inspect conditions
- name
- DeepSWE v1.1
- value
- 67.2
- unit
- %
- source Url
- https://openai.com/index/gpt-5-6/
- checked At
- 2026-09-08T19:30:56Z
- attribution
- provider-reported
- configuration
- OpenAI launch table; single-model column, not Ultra. Exact effort and harness not stated in the table.
Source reports, measured results, and editorial assessments are separate evidence types. Source dates do not imply a new benchmark run. Capture dates are shown when recorded; otherwise we say the capture date was not recorded. Recorded dates are compared with this catalogue snapshot (Sep 11, 2026), not today’s date.
Method AIS-V2
How this page is assessed
Quality claims follow Method AIS-V2.
Money never buys rank — commercial relationships cannot change rank, eligibility, evidence quality, or publication timing.
Submit or claim a listing is free and does not skip the editorial queue.
Private · this device only