DeepSeek Announced Aug 21, 2026

DeepSeek-V4-Flash-Vision-Exp

DeepSeek's first experimental multimodal V4 model, adding visual modules and continued training to V4-Flash at V4-Flash pricing, with publisher-reported agent results DeepSeek places close to Claude Opus 4.8.

Context
1M
1,048,576 positions per the published config
Off-peak discount
50% off peak
Peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays; cache hits bill $0.014 at peak
Vision input
Up to 600 images
JPEG, PNG, GIF, or WebP in user messages, up to 384 tokens each
Weights
Released
MIT license on Hugging Face; about 305B parameters

Publisher pricing

Pricing below comes from the release announcement and does not represent a MiniRouter route or price.

Publisher-listed API price phases in US dollars per million tokens.
PhaseInput / 1MOutput / 1MBlended / 1MEffectiveSource
Standard · peak$0.44$1.32$0.66Aug 21, 2026open-endedDeepSeek

Publisher-reported benchmarks

Results are grouped by unit. Percentage, Elo, and task-count results never share a scale.

BenchmarkReported resultReported byAs ofSource
Terminal Bench 2.183.9DeepSeekAug 21, 2026Announcement ↗
DeepSWE59.3DeepSeekAug 21, 2026Announcement ↗
CyberGym75.3DeepSeekAug 21, 2026Announcement ↗
NL2Repo57.7DeepSeekAug 21, 2026Announcement ↗
Toolathlon Verified75.9DeepSeekAug 21, 2026Announcement ↗
DSBench-Hard63.6DeepSeekAug 21, 2026Announcement ↗
ApexBench · pass@136.5DeepSeekAug 21, 2026Announcement ↗
Agents' Last Exam27.3DeepSeekAug 21, 2026Announcement ↗
Chartography64.3DeepSeekAug 21, 2026Announcement ↗
  • DeepSeek says the model matches V4-Flash on text agents, reasoning, and knowledge while bringing multimodal agent performance close to Opus 4.8.
  • DeepSeek labels the release experimental; the earlier non-vision V4-Flash ignored image content in prompts.
  • DeepSeek reports the benchmark figures; MiniRouter has not reproduced them.

Independent measurements

Artificial Analysis runs its own evaluations and timing against each model's first-party API. These are the only numbers on this page not reported by the publisher.

DeepSeek V4 Flash Vision (reasoning max) ranks #30 of 448 model families on their Intelligence Index.

Intelligence Index
51.5
Composite of their evaluations
Coding Index
65.0
Coding evaluations only
Blended price
$0.66
USD per 1M tokens, 3:1 input to output, first-party list
Output speed
111
Median tokens per second
First answer token
19s
Median seconds, including reasoning

DeepSeek V4 Flash Vision (reasoning max) on Artificial Analysis ↗

Source: Artificial Analysis, read Sep 2, 2026. Measured on the model's first-party API, not a MiniRouter route.

What is not independently confirmed

  • Every figure on this page is self-reported by the cited publisher. MiniRouter did not run these benchmarks.
  • Benchmark harnesses are not standardised. Results sharing a name across vendors have not been confirmed to use identical runs.
  • Publisher pricing can change, and introductory phases have stated end dates.
  • DeepSeek-V4-Flash-Vision-Exp is routable through MiniRouter today; live pricing on its model page is authoritative over the publisher figures here.

Related coverage

Other release dossiers and the current MiniRouter catalog.

Sources

  1. DeepSeek model card

    Published Aug 21, 2026

    Read at DeepSeek
  2. Artificial Analysis model page

    Published Sep 2, 2026

    Read at Artificial Analysis

Editorial reference revised 2026-09-02. Catalog status comes from the generated routing snapshot.