← Back to directory
OpenAI
GPT-4 Turbo
Released Nov 6, 2023 · knowledge cutoff 2023-04
- Status
- Deprecated
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 128K
- Max output
- 4K
- Speed
- —
- Price ($/MTok in / out)
- $10 / $30
- Cost per task
- —
- Open weights
- No
Benchmarks3 of 10 reported
MMLU-Proretired
63.7%
GPQA Diamond
48%
LMArena Elo
1256
SWE-bench Verified
—
SWE-bench Pro
—
Terminal-Bench 2.1
—
AIME (latest)retired
—
Humanity's Last Exam
—
GDPval-AA v2
—
ARC-AGI-2
—
Strengths & weaknesses
Strengths
- Noticeably quicker to respond than GPT-4 while keeping most of its judgment
- First OpenAI model that could hold a whole document or small codebase in view at once
- Reads screenshots, charts and scanned pages well enough to skip a separate OCR step
Weaknesses
- Truncates long code with placeholder comments instead of finishing the work — the 'lazy model' complaints
- Recall thins out toward the middle of a long prompt, so the bigger window isn't fully usable
- No way to make it think harder on a question it gets wrong
Announced at OpenAI's first DevDay. Knowledge cutoff corrected on 2026-09-03 to April 2023 per OpenAI's DevDay post for gpt-4-1106-preview; the December 2023 cutoff previously recorded belongs to the later gpt-4-0125-preview / gpt-4-turbo-2024-04-09 checkpoints, which the MMLU-Pro and GPQA figures here most likely measure.
For developers
API model strings
Not researched
Licence
Proprietary — weights not released
Retirement
Retirement date not researched
Lineage
No recorded predecessor