← DeepSeek V4 Pro · API · Max effort
Evidenceacceptedindependent_lab
GPQA Diamond
Artificial Analysis — GPQA Diamond (DeepSeek V4 Pro max)
ev_2026_07_dsv4pro_aa_gpqa
Result
88.8%
Confidence A
- Tested
- 2026-07-11
- Runs
- 1
Configuration
How it was measured
- Exact model
- deepseek-v4-pro
- Reasoning
- max
- Agent harness
- none
- Tool access
- none (closed-book scientific QA)
- Prompting
- Artificial Analysis standard GPQA Diamond protocol
PRIMARY INT — independent Artificial Analysis GPQA Diamond under max effort. Float locked via AA-sourced OpenRouter model card (AA does not expose raw % in static HTML; OpenRouter attributes metrics to AA).
Caveats
Known limitations
- GPQA near-saturated at frontier; use peer comparison not absolute %.
- Float transcribed from AA-sourced OpenRouter mirror of AA evaluation suite.
- Max effort — not default chat / Flash.
- Prefer this AA figure over vendor 90.1% for independent INT primary.
Weighting
Evidence weight factors
The engine weights each evidence row by these factors (0–1) when fusing scores — higher is more trusted.
relevance0.95
quality1.00
recency1.00
health0.70
comparability0.90
reliability0.85
Benchmark
GPQA Diamond
Graduate-level science questions (biology, physics, chemistry) designed to be Google-proof; Diamond is the hardest subset where PhD experts score ~65%.
Benchmark source →