TL;DR — Key Takeaways
- DeepSeek’s experimental V4-Flash-Vision-Exp model adds multimodal capabilities, including support for screenshots, charts and hundreds of images in a single request.
- DeepSeek says the model approaches Anthropic’s Opus-4.8 on several agent and multimodal benchmarks, although it still trails significantly on some coding tasks.
- The model’s biggest competitive advantage may be cost, with DeepSeek continuing to position its models as cheaper alternatives to high-end Western AI systems.
DeepSeek released an experimental multimodal model on Friday, claiming its agent capabilities rival those of Anthropic’s supported Opus-4.8 system.
The release marks the latest salvo from Chinese artificial intelligence (AI) firms seeking to close the gap with foreign rivals while offering drastically lower API pricing.
The newly deployed model, named DeepSeek-V4-Flash-Vision-Exp, builds upon the text-only V4-Flash architecture by enabling the system to interpret screenshots, charts, and up to 600 images in a single request. Alongside the model launch, the company released version 0.1.1 of its open-source agent harness to support visual processing tasks.
In self-published evaluations, DeepSeek pitted the vision model against Anthropic’s Opus-4.8 across 11 standardized tests. The experimental model edged out the American benchmark on three evaluations, leading by 1.3 points on DeepSWE, 1.6 points on Agents’ Last Exam, and 1.0 point on ZeroBench.
On eight other evaluations, the Chinese model trailed. While several scores remained within a fraction of a point — such as Toolathlon-Verified, where the models split 75.9 to 76.2 — DeepSeek fell sharply behind on repository-level code tasks. On NL2Repo, a core benchmark for corporate software workflows, Opus-4.8 beat DeepSeek by 12 points, 69.7 to 57.7.
The vendor’s disclosure also revealed that gains over its previous text-only variant were partially structural. In multimodal comparisons, the older V4-Flash was scored on tasks containing images it could not process, making the visual model’s jump as much a reflection of newfound sight as optimized logic. However, adding visual components introduced a slight performance trade-off in security tests, where the vision variant dropped 1.4 points against its predecessor on Cybergym.
Notably, DeepSeek chose to measure its platform against Opus-4.8, a model active through at least May 2027, rather than Anthropic’s newer Claude Opus 5 flagship. DeepSeek made no claims regarding performance against Opus 5, leaving direct comparison data between the top-tier systems unverified by independent reviewers.
Despite minor performance gaps on major tasks, DeepSeek’s release highlights a growing commercial divide in the artificial intelligence market: raw efficiency versus operating cost.
Industry figures show V4-Flash operates as one of the most cost-effective foundation models available, processing one million text words for roughly 87 cents compared to approximately $50 for high-end rival services. For enterprise buyers considering routine agent deployments, a slight difference in general benchmark scores is increasingly weighed against significant operational savings.
DeepSeek simultaneously confirmed the official release of V4-Pro, a production-grade model featuring integrated Codex support designed for heavy enterprise workloads.
The launch reflects broader momentum across China’s AI sector, where developers are rapidly expanding infrastructure and multimodal features. As global geopolitical tensions prompt stricter regulatory scrutiny over international technology supply chains, Chinese firms continue pushing competitive hardware utilization to capture developer adoption worldwide.
