Preview DeepSeek V4 Flash checkpoint. Reasoning effort modes remain variants of this revision.
Human Frontier
85.8
Human-calibrated frontier signal, backed by Arena-style preference evidence and separate from raw AgMoBench benchmark composite scores.
Blended Price
$0.17/M
Input Price
$0.13/M
Output Price
$0.28/M
Speed
—
TTFT
—
Axes
Filters
Providers
vcontinuous-leaderboard · tasks-created:2026-03-01:2026-05-15 · DeepSeek-V4 Flash · harness swe-rebench-react@tools · 5 attempts · pass@1 · swe-rebench:standardized:128k-or-model-limit · provider DeepSeek-V4 Flash__tools · record DeepSeek-V4 Flash__tools:1772323200000:1778803200000
{
"passAt5": 57.27272727272727,
"developer": "deepseek",
"instanceType": "model",
"contextPolicy": "128k unless the model supports less",
"attemptsPerTask": 5,
"totalTokenUsage": 2996077.38,
"modelReleaseDate": "2026-04-24",
"scoreStandardError": 0.9706434774573924,
"cachedTokenPercentage": 95.49776510326375,
"modelReleaseTimestamp": 1776988800000,
"standardContextTokens": 128000,
"averageInstanceCostUsd": 0.06918754347272726,
"potentialContamination": true,
"aggregateTaskRangeTimestamp": {
"to": 1778803200000,
"from": 1772323200000
},
"publishedTaskRangeTimestamp": {
"to": 1778803200000,
"from": 1772323200000
}
}Evidence observed 2026-05-15