DeepSeek officially released deepseek-v4-flash-vision-exp on August 21, 2026, as reported by The Decoder, expanding its V4-Flash family with multimodal input. The experimental model keeps everything the text version does — reasoning, world knowledge, language ability — and adds understanding of images and screenshots. In practice, that means it can look at a UI, read a chart, or inspect a screenshot as part of an agent workflow, not just process text.
DeepSeek also updated DeepSeek Harness, its agent software, to version 0.1.1 with native out-of-the-box support for the new model, and launched a Files API alongside it. The package is coherent: a model that can see, a harness that can act, and an API for file-based workflows.
The benchmark picture: close to Opus 4.8, but not a clean win
DeepSeek's evaluation data compares the vision model with Anthropic's Claude Opus 4.8 across 11 multimodal agent benchmarks. The result is a draw with an edge: DeepSeek wins 3, loses several, and comes close on the rest. The wins are not trivial:
| Benchmark | DeepSeek vs. Claude Opus 4.8 |
|---|---|
| DeepSWE (software engineering agents) | +1.3 points for DeepSeek |
| Agents' Last Exam (agentic knowledge tasks) | +1.6 points for DeepSeek |
| ZeroBench (hard visual reasoning) | +1.0 point for DeepSeek |
| NL2Repo (natural language to repository) | 57.7 vs. 69.7 — Opus 4.8 leads by 12 |
Beating Opus 4.8 on DeepSWE, a demanding software-engineering agent test, and on ZeroBench, which was designed to break visual models, is genuinely notable for a model at this price point. But the 12-point loss on NL2Repo is a reminder that "rivals Opus 4.8" does not mean "replaces Opus 4.8." Also interesting: on the earlier Terminal Bench 2.1, the text-only V4-Flash scored 82.7 against 72.1 for DeepSeek's own V4-Pro-Preview — a sign that the Flash family is unusually strong at agentic terminal work.
The price: this is where it gets uncomfortable for US labs
DeepSeek kept the standard V4-Flash pricing for the vision variant: $0.14 per 1M input tokens and $0.28 per 1M output tokens — roughly €0.13 and €0.26 at current exchange rates, per DeepSeek's API platform. Images are tokenized at up to 384 tokens per image. Do the math: one screenshot costs about $0.000054 in input tokens. A batch of 10,000 images runs roughly $0.54. That is not a rounding error; that is a rounding error's rounding error.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| DeepSeek V4-Flash-Vision-Exp | $0.14 (≈€0.13) | $0.28 (≈€0.26) |
| GPT-5.6 (Sol tier) | $0.20 | $1.20 |
| Gemini 3.7 Flash | $0.75 | $3.75 |
| Grok 4.6 | $2.00 | $6.00 |
| Llama 4 (open weights) | $0 | $0 |
| Mistral Leanstral 1.5 / Medium 3.5 | $0 self-hosted (Apache 2.0) | $0 self-hosted |
The model also brings a 1,000,000-token context window and up to 384,000 output tokens — more context than most hosted rivals at any price, let alone this one.
The European angle: available, but know your GDPR exposure
Is the model available in the EU? Yes — DeepSeek's API is accessible from Europe, and pricing is the same for EU users. The catch is not availability; it's data location. An API call to DeepSeek means data is processed on infrastructure outside the EU, which triggers GDPR Chapter V transfer rules for any personal data. A standard contractual clause is the minimum; a data-protection impact assessment may be required for sensitive workloads.
The EU AI Act context matters too. The Digital Omnibus on AI (Regulation (EU) 2026/1744) postponed the Annex III high-risk obligations to December 2, 2027, but GPAI transparency obligations under Article 50 are already active — so DeepSeek, as a general-purpose AI provider, operates under enforceable EU rules, not a regulatory vacuum. European teams that want vision-capable agents without cross-border transfers can use open-weight Llama 4 (natively multimodal) or Mistral's Leanstral 1.5 (Apache 2.0, text-oriented) on their own hardware.
What to do with it
Concretely, this model is interesting for UI screenshot QA and visual regression testing, document workflows that include figures and charts, and — most importantly — cheap agent experiments that need a "seeing" step. A developer can now build a screenshot-driven automation loop for a fraction of the cost of the big US APIs.
Treat the -exp suffix seriously. Experimental models change behavior without notice; pin versions, keep fallbacks, and do not put this in a production pipeline without a robustness layer. On our AI Arena rig we benchmark exactly this class of local multimodal models, and open-weight alternatives are improving faster than most people realize — but for hosted vision agents at this price, DeepSeek just moved the goalposts.
Is deepseek-v4-flash-vision-exp open-source?
No. It is API-only and explicitly experimental. DeepSeek has not announced open weights for the V4-Flash family. If you need fully on-prem vision, Meta's Llama 4 is the closest free multimodal alternative; Mistral's Leanstral 1.5 is Apache 2.0 but text-only.
How much does one image actually cost in API tokens?
Images are tokenized at up to 384 tokens each. At $0.14 per 1M input tokens, that is about $0.000054 (≈€0.00005) per image in input tokens, plus whatever text tokens the conversation uses. A 10,000-image batch runs roughly $0.54.
Why does DeepSeek compare against Opus 4.8 and not Claude Opus 5?
Because that is the comparison DeepSeek published. Anthropic released Claude Opus 5 on July 24, 2026, but no official head-to-head between Opus 5 and DeepSeek's vision model has been published yet.
