On August 21, DeepSeek launched 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 is not a routine minor update—it packs "vision agent" capabilities, previously reserved for premium frontier tiers, into the cheapest frontier-class model at just $0.14 per million input tokens.
The release itself
The new model retains all V4-Flash text capabilities (reasoning, world knowledge, agent behavior) while adding image and screenshot understanding. According to DeepSeek, on 11 multimodal agent benchmarks its performance is close to Anthropic's Claude Opus 4.8:
- DeepSWE (software engineering agent): leads by 1.3 points
- Agents' Last Exam: leads by 1.6 points
- ZeroBench (a hard benchmark designed to break vision models): leads by 1.0 points
- NL2Repo (natural language to repo): trails 57.7 vs 69.7, a 12-point gap
- Images billed at up to 384 tokens each, same rate as text (~$0.000054 per screenshot; ~$0.54 for 10,000 images)
- 1M-token context window
- Up to 600 images per request
- The experimental tag means no API stability guarantees; production systems should pin expectations carefully.
- Data leaves China (processed on infrastructure outside China); for EU users this triggers GDPR Chapter V transfer rules—enterprise users must assess compliance themselves.
Pricing details:
Harness 0.1.1 accompanies the model with key workflow support: /goal and /plan commands accept image-text input, the @ menu references files and sessions, MCP/ACP supports persistent image attachments, and PTC Mode forwards nested images. A free Files API lets you upload once and reference by file_id across requests, eliminating repeated base64 encoding. The model is compatible with Chat Completions, Messages, and Responses endpoints, so it can drop into existing agent frameworks directly.
Why it matters
1. Multimodality is becoming a commodity. Six months ago, vision models were a separate, expensive tier. DeepSeek now bundles vision into a budget model at text-equivalent pricing—screen reading, document processing, and visual QA workflows no longer pay a frontier-vision premium. This directly rewrites the ledger for agent builders.
2. An agent's "eyes" and "hands" are decoupling in pricing. The model sees (vision understanding) + Harness executes (tool calls) + Files API manages assets—a coherent package that makes the "screenshot in, action out" loop viable at extremely low cost.
3. Competitive pressure spreads. GPT-5.6 Sol sits at $0.20, Gemini 3.7 Flash at $0.75, and Grok 4.6 at $2.00 per million input tokens. DeepSeek has pushed the entry price for vision agents down by an order of magnitude, forcing frontier labs to rethink what vision should cost.
Caveats
*Sources: DeepSeek official X announcements and API documentation, The Decoder, Times of India, Beijing News Shell Finance, ai-jarvis.eu benchmark analysis.*