Support

DeepSeek · released 2026-08-21

DeepSeek V4 Flash Vision

Measured on 34 benchmarks.

Compare

24.3

#52 of 109 models · best at max

includes estimates: 45% of the mix is measured for this model

Strengths

  • DeepSeek V4 Flash Vision
  • Fable 5.2 (#1)
  • Field median
bestField medianFrontierCode 2.0-AATerminal-Benchv4-AAAutomationBench-AAAA-Briefcase v1.1GDP.pdfTerminal-Bench-Science-AACritPt v2 HardSubsetAA-Omnisciencev1.1 AccuracyHLE-DiamondAA-Omnisciencev1.1Non-hallucination
  1. FrontierCode 2.0-AA15%17.9% est.#45 of 109
  2. Terminal-Bench v4-AA10%12.1%#53 of 109
  3. AutomationBench-AA5%47.5%#56 of 109
  4. AA-Briefcase v1.115%1418#42 of 109
  5. GDP.pdf5%12.0%#62 of 109
  6. Terminal-Bench-Science-AA15%4.5% est.#48 of 109
  7. CritPt v2 Hard Subset10%27.2% est.#54 of 109
  8. AA-Omniscience v1.1 Accuracy10%34.0% est.#51 of 109
  9. HLE-Diamond10%21.1%#42 of 109
  10. AA-Omniscience v1.1 Non-hallucination5%32.4% est.#88 of 109

Bars show where DeepSeek V4 Flash Vision sits among the 109 ranked models on each benchmark (right is best); the tick is the field median.

By category

  • DeepSeek V4 Flash Vision
  • Fable 5.2 (#1)
  • Field median
p25p50p75topField medianCodingAgents & computeruseProfessional workScience & researchReasoning & puzzlesKnowledge &accuracyVisionLong documents
  1. Coding5 benchmarks39%#49 of 106
  2. Agents & computer use10 benchmarks41%#48 of 109
  3. Professional work9 benchmarks47%#44 of 109
  4. Science & research4 benchmarks56%#25 of 105
  5. Reasoning & puzzles5 benchmarks47%#40 of 101
  6. Knowledge & accuracy3 benchmarks47%#46 of 102
  7. Vision7 benchmarks24%#67 of 88
  8. Long documents3 benchmarks44%#52 of 101

Across 46 results on every current benchmark that measured DeepSeek V4 Flash Vision, not only this ranking’s. The percentage is its average standing among the 109 models in this ranking that took the same benchmarks: 100% beat them all, 0% trailed them all. A benchmark in two categories counts in both; a category needs two benchmarks.

Cost and speed

against 109 ranked models

Output speed

#24 fastest of 84

160 tok/sat max

Time to first token

#18 quickest of 84

1.8sat max

Token price

$0.031 per 1M tokens at API prices

97.0% cache read · 2.50% fresh input · 0.00% cache write · 0.50% output

Assumed mix: no measured T3 Code usage for this model.

Not measured for DeepSeek V4 Flash Vision: cost per task, time per task, output tokens per task (no published API price yet).

Availability

API providers

3
ProviderIn / out per 1MTok/sFirst
  • DeepInfra$0.22 / $0.65851.8s
  • GMI$0.44 / $1.321602.3s
  • SiliconFlow$0.44 / $1.321651.2s

OpenRouter, Artificial Analysis

This ranking’s benchmarks

5 of 10 measured

Cost per task is what a subscriber pays: the model’s plan price, a share of API price. Models no plan sells are at API price.

Agents & computer use

AutomationBench-AA

283 results
  1. #133max of 28347.5%

Leader: Gemini 4 Neon · 79.3%

Professional work

AA-Briefcase v1.1

288 results
  1. #86max of 2881418

Leader: Sonnet 5.5 · 1813

Professional work

GDP.pdf

283 results
  1. #164max of 28312.0%

Leader: Gemini 4 Neon · 33.4%

Knowledge & accuracy

HLE-Diamond

197 results
  1. #91max of 19721.1%

Leader: Fable 5.2 · 80.2%

Agents & computer use

Terminal-Bench v4-AA

280 results
  1. #129max of 28012.1%

Leader: Kimi K3.1 · 67.2%

All 34 benchmarks

show 29 more