DeepSeekChina· released Aug 2026· open weights

DeepSeek V4 Pro

Name: DeepSeek V4 Pro
Author: DeepSeek

Ranked #20, 32.5% behind the frontier on average.

DeepSeek, China. One of 4 models from this lab on the Index.

Third-party publishedMedium confidence (7+ comparable tests, 3+ Index categories)

Available in Theo as Theo Open Pro

Theo orchestrates it for the steps it does best and always shows which engine answered.

Use it in Theo

Snapshot September 8, 2026

OpenCharts Index

67.5

Ranked #20 (17 to 23) of 36 · 32.5% behind the frontier on average · 16 comparable tests · 3 counting categories

Drop any single test and the Index lands between 63.3 and 71.2, anywhere from #17 to #23. Neighbours inside that range are ties.

4 of 8 hardest tests · share of the frontier

Distance to today's frontier on the hardest tests. Not a measure of intelligence.

What it's known for

DeepSeek V4 Pro, in 16 results.

Strongest in reasoning (76.6) and coding (64.6). It trails the frontier most in math, 38.7% behind on average. No comparable results yet in multimodal and long context. Between 30% and 35% behind the best published results, on average. Drop any single test and it would sit anywhere from #17 to #23.

Strongest category

Reasoning

Hard, novel problems: graduate science, abstract puzzles, expert exams.

76.6

GPQA Diamond91.7% · 94% of best
ARC-AGI-190.5% · 92% of best
ARC-AGI-261.3% · 65% of best
CritPt18.0% · 56% of best

Mean of 4 of the category's 6 comparable tests. Counts toward the Index.

Best single result

Mock AIME 2024–2025

Solve competition mathematics problems that have a single exact integer answer.

98.6%

98.6% of the best (100.0%) · #9 of 78.

Olympiad-style problems need long, careful chains of reasoning, and there is no partial credit.

Where it trails

FrontierMath Tier 4

Solve the hardest FrontierMath tier: problems that take expert mathematicians days.

26.8%

72.5% behind the best published result (97.6%) · #25 of 51.

The deepest end of the set. Progress here signals genuinely new capability.

Not measured yet

No comparable result yet in these 2 categories. A blank is a blank, never a zero, and it does not lower the Index.

Multimodal
Long context

Covered, but not averaged into the Index:

Knowledgesingle-test category, shown beside the Index · 70.0
Agentic1 of 8 tests · needs 4 · mean 60.6
Human preferencesingle-test category, shown beside the Index · 86.4

Every score, with its source

See it at work

What DeepSeek V4 Pro was asked to do, and how it did.

Every test it has taken, by category, with the kind of task it faced, who produced the number and where the result landed against the best published one. The examples are original and illustrative, never items from the datasets themselves.

Human preference · mean 86.4 across 1 comparable test · a single-test category, shown beside the Index and never averaged into it

Text ArenaLeaderboard

Answer the same prompt as a rival model; real users vote blind on which answer they prefer.

1,460

43% expected win rate against the board leader (1,507)
#39 of 92

high-20260813 · board 2026-09-02

Source: Arena

Every result is read against the best published one on its test: percent scores as a share of the best above chance, Arena ratings as the expected win rate against the board leader, open-ended values on a log scale. Sitting a harder exam never lowers a model. Every number links to the publisher that produced it.

Head to head

DeepSeek V4 Pro against whoever you pick.

Choose a rival. Every counted test both have taken appears side by side as a share of the best published result on that test, with the category means above. A win is a higher share; a gap under 2 points is a tie, because one item on a 45-problem exam is 2.2 points and publishers' own standard errors are of that order.

DeepSeek V4 Pro

#20·Index 67.5C+

0 : 12

wins · 1 tie · 13 shared

GPT-6 Astra

#1·Index 99.0A+

GPT-6 Astra comes out ahead on 12 of the 13 tests both have taken.

Category means

Reasoning
76.6
99.6
Knowledge
70.0
100.0
Coding
64.6
97.7
Math
61.3
99.8
Agentic
60.6
98.5
Human preference
86.4
—

Shared tests · biggest gaps first

FrontierMath Tier 4Math
26.8%
97.6%
−72.5
FrontierCodeCoding
17.6%
53.3%
−66.6
Code Arena (WebDev)Coding
1,582
1,797
−55.1
ProofBenchMath
50.0%
99.0%
−49.0
CritPtReasoning
18.0%
31.7%
−42.5
ARC-AGI-2Reasoning
61.3%
95.0%
−35.5
FrontierMath (Tiers 1–3)Math
64.6%
93.7%
−31.1
SimpleQA VerifiedKnowledge
52.9%
75.6%
−30.0
WeirdMLCoding
66.2%
92.9%
−28.7
SciCodeCoding
50.0%
56.5%
−10.4
ARC-AGI-1Reasoning
90.5%
98.5%
−8.1
GPQA DiamondReasoning
91.7%
95.8%
−5.8
Mock AIME 2024–2025Math
98.6%
100.0%
tie

Bars are distances to the best published result on each test (percent scores above chance, Arena ratings as win rate against the leader, open-ended values on a log scale). Raw values as the publishers report them.

Open GPT-6 Astra

Scores

Every test, every source

The best published run per benchmark, the raw value as the publisher reports it, who produced the number, how close it comes to the best published result on that test, and where the model sits among every model scored on it. Percent scores are read above chance, Arena ratings as a win rate against the board leader, open-ended values on a log scale. A test with too few models to compare against is shown but not counted.

Category means

Reasoning76.6 · 4/6
Knowledgebeside70.0 · 1/1
Coding64.6 · 5/7
Math61.3 · 4/5
Agentic60.6 · 1/8
Multimodal—
Human preferencebeside86.4 · 1/1
Long context—

Mean · tests taken of the category's comparable tests. A faded category either rests on fewer than half of them or is a single-test category (marked “beside”), which is shown next to the Index and never averaged into it.

GPQA DiamondEpoch-runReasoning

91.7%

94.2% of the best (95.8%), above chance
79th percentile of 78 · n = 198

max

Source: Epoch AI — AI Benchmarking Hub

ARC-AGI-2LeaderboardReasoning

61.3%

64.5% of the best (95.0%)
60th percentile of 51 · n = 120

max

Source: Epoch AI — AI Benchmarking Hub

ARC-AGI-1LeaderboardReasoning

90.5%