DeepSeek Gives Its Cheapest Frontier Model Eyes: V4-Flash-Vision-Exp + Harness 0.1.1 Released Same Day
On August 21, DeepSeek put the experimental multimodal model deepseek-v4-flash-vision-exp on its API platform, and released DeepSeek Harness 0.1.1 the same day with out-of-the-box support for the new model. This isn't a routine minor update: it packs "vision agent" capabilities—previously reserved for expensive frontier tiers—into the cheapest frontier-class model at just $0.14 per million input tokens.
What happened
The new model keeps all of V4-Flash's text capabilities (reasoning, world knowledge, agent behavior) and adds image and screenshot understanding. According to DeepSeek, across 11 multimodal agent benchmarks it performs close to Anthropic's Claude Opus 4.8:- DeepSWE (software engineering agent): ahead by 1.3 points
- Agents' Last Exam: ahead by 1.6 points
- ZeroBench (a hard benchmark designed to defeat vision models): ahead by 1.0 points
- NL2Repo (natural language to repo): behind, 57.7 vs 69.7 — a 12-point gap
Harness 0.1.1 matters just as much: /goal and /plan commands accept image+text input, the @ menu can reference files and sessions, MCP/ACP support persists image attachments, and PTC Mode forwards nested images. A free Files API also launched—upload once, reference by file_id across requests, no repeated base64. The model is compatible with the Chat Completions, Messages, and Responses endpoints, so it slots directly into existing agent frameworks.
Why it matters
1. Multimodal is dropping from the premium tier to the commodity tier. Six months ago vision models were a separate, expensive tier; now DeepSeek bundles vision into a budget model at text-equivalent pricing. Workflows like screen reading, document processing, and visual QA no longer require paying a frontier-vision premium—an immediate rewrite of agent builders' budgets.
2. Pricing for an agent's "eyes" and "hands" is decoupling. A model that can read images (vision) + Harness for execution (tool calls) + Files API for assets form a coherent package, meaning a "screenshot in, action out" loop can now run at extremely low cost.
3. Competitive pressure spills over. Compared to GPT-5.6 Sol at $0.20, Gemini 3.7 Flash at $0.75, and Grok 4.6 at $2.00 per million input tokens, DeepSeek pushes the entry price for vision agents down by an order of magnitude, forcing frontier labs to rethink what vision should cost.
Caveats
The experimental tag means no API stability guarantee—production systems should pin expectations carefully. Data residency: the API processes requests on infrastructure outside China; for the EU this triggers GDPR Chapter V transfer rules, so enterprise users must assess compliance themselves.One line: when the cheapest model can read your screen, the inflection point for democratized agents has arrived.
---
Sources: DeepSeek's official X announcements and API docs, The Decoder, Times of India, Beijing News Shell Finance, and benchmark breakdowns from ai-jarvis.eu.