DeepSeek V4 Pro

AA lists DeepSeek V4 Pro as a top-three open-weights model with Intelligence Index 52.

66 Floor score

confidence: medium

66 Floor Capability

Finished benchmark shape.

72 Operating Envelope

Lower DSWE pass rate, but cost, time, and output tokens are contained.

71 Frontier Reality

AA open top-3 / Index 52

62 Research Trust

Current open-model baseline; source-discipline evidence is not yet deep enough.

Current signal

AA open top-3 / Index 52

Evidence completeness

54%

4 score dimensions represented.

DSWE

8% pass@1

$4.22 / 37m / 50k output tokens

Calculation

Score components

64 Accuracy + grounding

Correct, grounded answers on ordinary verifiable tasks.

66 Instruction discipline

Follows exact user constraints and output formats.

62 Honesty + reliability

Avoids unsupported confidence and remains stable across runs.

70 Reasoning floor

Handles non-flashy reasoning without brittle failures.

70 Agent/coding

Completes tool, coding, and agent-style tasks when evidence exists.

70 Operating envelope

Cost, speed, token burn, verbosity, and friction.

Evidence

Sources used

Artificial Analysis open-weights leaderboard
Arena/Kilo coding signals

Caveat

What to watch

Included as a current open-model baseline. Needs fresh internal floor prompts before it can be trusted as a user-facing assistant score.

Meaning

How to read the score

Each score is a 0-100 saturation scale. A score of 100 means that benchmark lane is complete for this model. Floor Capability is the finished benchmark shape; Operating Envelope, Frontier Reality, and Research Trust are provisional tabs that need prompt-pack backfill. Confidence is evidence quality, not another model grade. Evidence % is source coverage, not capability.

Citations

Benchmark sources

Artificial Analysis open-weights leaderboard