Text ArenaLeaderboard
Answer the same prompt as a rival model; real users vote blind on which answer they prefer.
1,418
37% expected win rate against the board leader (1,507)
#74 of 92
thinking · board 2026-09-02
Source: ArenaDeepSeekChina· released Sep 2025· open weights
DeepSeek V3.1 Terminus
Provisional: 3 comparable tests so far. Ranked once it reaches 6 across 3 counting categories.
DeepSeek, China. One of 4 models from this lab on the Index.
Third-party publishedLow confidence (thin coverage)
Snapshot September 8, 2026
OpenCharts Index
48.5
Provisional: not enough comparable tests or counting categories to rank yet · 3 comparable tests · 0 counting categories
Drop any single test and the Index lands between 35.4 and 70.2.
Needs more results (0 of 8 so far)
Distance to today's frontier on the hardest tests. Not a measure of intelligence.
What it's known for
DeepSeek V3.1 Terminus, in 3 results.
Its widest gap is on CritPt, 94.7% behind the best published result. No comparable results yet in knowledge, math, agentic, multimodal and long context. Between 50% and 60% behind the best published results, on average.
Strongest category
No category counts yet: a category counts once it has more than one comparable test and the model has taken at least half of them. The scores it has are below.
Best single result
Text Arena
Answer the same prompt as a rival model; real users vote blind on which answer they prefer.
1,418
37% expected win rate against the board leader (1,507) · #74 of 92.
Millions of blind votes on real prompts are the market's measure of what people prefer.
Where it trails
CritPt
Solve an unpublished research-level physics problem to a numeric or symbolic answer checked by an official server.
1.7%
94.7% behind the best published result (32.3%) · #57 of 78.
Research physics is far past textbook recall. Only a few models produce anything a physicist would accept.
Not measured yet
No comparable result yet in these 5 categories. A blank is a blank, never a zero, and it does not lower the Index.
- Knowledge
- Math
- Agentic
- Multimodal
- Long context
Covered, but not averaged into the Index:
- Reasoning1 of 6 tests · needs 3 · mean 5.3
- Coding1 of 7 tests · needs 4 · mean 65.5
- Human preferencesingle-test category, shown beside the Index · 74.8
See it at work
What DeepSeek V3.1 Terminus 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 74.8 across 1 comparable test · a single-test category, shown beside the Index and never averaged into it
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 V3.1 Terminus 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 V3.1 Terminus
provisional·Index 48.5D
0 : 3
wins · 0 ties · 3 shared
DeepSeek V4 Pro
#20·Index 67.5C+
DeepSeek V4 Pro comes out ahead on 3 of the 3 tests both have taken.
Category means
Reasoning
5.376.6Knowledge
—70.0Coding
65.564.6Math
—61.3Agentic
—60.6Human preference
74.886.4
Shared tests · biggest gaps first
- CritPtReasoning−50.41.7%18.0%
- SciCodeCoding−15.140.6%50.0%
- Text ArenaHuman preference−11.61,4181,460
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 DeepSeek V4 ProScores
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
- Reasoning5.3 · 1/6
- Knowledge—
- Coding65.5 · 1/7
- Math—
- Agentic—
- Multimodal—
- Human preferencebeside74.8 · 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.
CritPtLeaderboardReasoning
1.7%
5.3% of the best (32.3%)
26th percentile of 78
SciCodeLeaderboardCoding
40.6%
65.5% of the best (62.0%)
21th percentile of 73 · n = 338
Text ArenaLeaderboardHuman preference
1,418
37% expected win rate against the board leader (1,507)
20th percentile of 92
thinking · board 2026-09-02
Source: Arena (LMArena) — Leaderboard DatasetTop of the leaderboard
Compare with its neighbours.
The best models, ranked here, working inside Theo.
28 of the models on this page run inside Theo today. Theo picks the right one for each step and always shows which engine answered.