PodcastIntelligence Snacks · Weekly conversations about AI, software and business, from the team behind Normie Mode.Listen →
Models · DeepSeek

DeepSeek V4 Pro

DeepSeek's DeepSeek model, released 13 Aug 2026.

Top three for

No task yet.

Evidence

Independently tested. 17 independent results so far.

Cost

$1.48 per 1,000 typical requests, if you use it directly.

Reads at once

1,500 pages of about 500 words.

Where it ranks, task by task

Its place on each task's score, among the models we compare that have enough results to score.

TaskPositionTask score
Best AI for Excel and spreadsheets5th of 865
Best AI for business7th of 838
Best AI for legal questions7th of 846
Best AI for accounting and finance7th of 838
Best AI for research8th of 1142
Best AI for documents, PDFs and notes8th of 1356
Best AI for presentations and PowerPoint8th of 834
Best AI for writing9th of 1435
Best AI for health questions9th of 1440
Best AI for planning and personal admin10th of 1441
Best AI for studying and homework10th of 1260
Best AI for translation and languages10th of 1440
Best AI for maths and science10th of 1263
Best AI for coding and building apps10th of 109
Best AI for CVs, job applications and interviews10th of 1433

Test results

Every test we use that has a result for this exact version.

TestResultPosition
Competition maths
Epoch AI
99%5th of 12
Pulling information out of documents into tables
DTBench authors
94%7th of 13
People's votes on creative writing
LMArena
rating 14388th of 14
People's votes on legal questions
LMArena
rating 14638th of 14
People's votes on health questions
LMArena
rating 14448th of 13
Graduate-level science questions
NYU and others
92%7th of 11
Answering short factual questions correctly
Google DeepMind
53%7th of 11
People's votes overall
LMArena
rating 14459th of 14
People's votes on science questions
LMArena
rating 14579th of 14
People's votes on writing and language
LMArena
rating 14439th of 14
People's votes on business and finance questions
LMArena
rating 144310th of 14
People's votes on back-and-forth conversations
LMArena
rating 145510th of 14
People's votes on questions in other languages
LMArena
rating 143110th of 14
Research-level maths
Epoch AI
65%8th of 11
Professional tasks in banking, consulting and law
Mercor
47%6th of 8
People's votes on coding questions
LMArena
rating 147011th of 14
People's votes on following instructions
LMArena
rating 144711th of 14
People's votes on long, detailed requests
LMArena
rating 145911th of 14
Answering business questions from a spreadsheetTested by us
Normie Mode
29/305th of 6
People's votes on maths questions
LMArena
rating 144111th of 13
People's votes on expert questions
LMArena
rating 145812th of 14
Would a real maintainer accept its code?
Cognition
29%9th of 9

Details

Maker
DeepSeek
Released
13 Aug 2026
Status
Current
Replaces
DeepSeek V3.2
Name checked
Confirmed with DeepSeek
Intelligence Snacks newsletter

The big AI ideas each week, in your inbox