← Back to directory
OpenAI
GPT-5
Released Aug 7, 2025 · knowledge cutoff 2024-10
- Status
- Superseded
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 400K
- Max output
- 128K
- Speed
- —
- Price ($/MTok in / out)
- $1.25 / $10
- Cost per task (high effort)
- $0.22
- Open weights
- No
Benchmarks8 of 10 reported
MMLU-Proretired
87%
GPQA Diamond
85.7%
SWE-bench Verified
74.9%
SWE-bench Pro
41.78%
AIME (latest)retired
94.6%
Humanity's Last Exam
24.8%
LMArena Elo
1442
ARC-AGI-2
9.9%
Terminal-Bench 2.1
—
GDPval-AA v2
—
Strengths & weaknesses
Strengths
- Decides for itself when a question deserves deliberation, so you don't pick a model per task
- Much readier to say it doesn't know instead of inventing an answer
- Holds up better across long multi-file coding sessions than any earlier OpenAI model
Weaknesses
- Routing misjudgments made quality feel random — the same prompt could get a shallow or a deep answer
- Flatter, more clinical voice; users described losing a companion, not just a tool
- Under-answers by default, giving a summary where the previous line elaborated
Scores shown for GPT-5 (thinking, high). SWE-bench Pro 41.78% is Scale's public-set leaderboard figure for "gpt-5-2025-08-07 (High)" — third-party, not OpenAI-reported; the original SWE-Bench Pro paper (arXiv 2509.16941) lists the same run at 41.8%, while the 23.3% widely quoted from the paper's launch coverage was the commercial-set figure.
For developers
API model strings
- openai
gpt-5-2025-08-07
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor