DeepSeek launches multimodal AI model nearing Anthropic Opus 4.8 performance at 99% lower cost

DeepSeek unveils vision model challenging Anthropic's Opus 4.8 performance

DeepSeek has launched an experimental multimodal model called DeepSeek-V4-Flash-Vision-Exp, which it claims approaches the performance of Anthropic's Claude Opus 4.8 on various benchmarks. This release signifies DeepSeek's continued efforts to enhance visual AI capabilities while maintaining lower costs compared to American competitors. The model was introduced on Friday, marking an important step in the company's strategy to compete in the rapidly evolving AI landscape. The new model is a multimodal variant of DeepSeek's existing V4-Flash text model. It preserves the original text functionalities, such as reasoning and general knowledge, while incorporating image understanding features. According to DeepSeek's benchmarks, the DeepSeek-V4-Flash-Vision-Exp model significantly narrows the performance gap with Opus 4.8. For instance, on the ApexBench, the new model scored 36.5 at Pass@1, while Opus 4.8 achieved 39.4. Moreover, on Agents' Last Exam, DeepSeek's model outperformed Anthropic's with a score of 27.3 compared to Opus 4.8's 25.7. The model also achieved a score of 35.0 on ZeroBench, surpassing Opus 4.8, which scored 34.0. Despite its promising results, DeepSeek's evaluation of its models was conducted using its internal Harness Minimal Mode, meaning that the performance figures have not been independently verified. The DeepSeek-V4-Flash-Vision-Exp model is now available on the DeepSeek API, allowing developers to integrate it into their applications. The pricing structure is designed to keep costs lower than comparable models from U.S. labs, with images billed by token count and capped at 384 tokens per image. In conjunction with the model release, DeepSeek also introduced version 0.1.1 of its Harness framework, which offers built-in support for the new model. Additionally, a Files API has been made available, enabling developers to upload images once and reference them multiple times using a file ID, without incurring costs. This launch occurs amid heightened competition from both domestic rivals like Alibaba and Moonshot AI, as well as U.S. firms such as Anthropic and OpenAI. Bloomberg has framed this release as part of DeepSeek's ongoing challenge to Anthropic's standing in advanced AI technology. The introduction of the DeepSeek-V4-Flash-Vision-Exp model reflects a broader trend among Chinese AI labs to enhance multimodal and agentic capabilities within mid-tier models. The future of this experimental model will depend on developer adoption, indicating that the next phase of AI competition will increasingly focus on affordable models that can perform visual reasoning tasks.

DeepSeek says its new multimodal model nears Anthropic's Opus 4.8: Here's how

DeepSeek just landed yet another punch in the arms race for AI's price/performance ratio - and this punch has eyes. Hangzhou-based company DeepSeek announced the release of the DeepSeek-V4-Flash-Vision-Exp - an experimental multimodal variation of its star performer, the V4 Flash. While DeepSeek-V4-Flash could only process textual information, DeepSeek-V4-Flash-Vision-Exp can now process images and screenshots, effectively giving its star performer eyes. Also read: Why right now is the best time to buy the Samsung Galaxy Z Fold 8 and Fold 8 Ultra The most noteworthy part of the release is the framing of it by DeepSeek itself. In the release statement, DeepSeek claims that its experimental AI model "matches DeepSeek-V4-Flash in text capabilities, including agents, reasoning, and world knowledge," and that when benchmarked as a multimodal agent, it "makes a major leap in terms of performance," reaching a performance level "close to Opus-4.8," the most advanced model from Anthropic's. The key words here are "close to," and it should be taken into account before any winner is proclaimed. DeepSeek made eleven comparisons with Opus 4.8. The model beats Opus in three of those, outclassing it in DeepSWE, Agents' Last Exam, and ZeroBench by about 1 to 1.6 points margin. In the remaining eight comparisons, it loses to Opus 4.8, and in at least one of the hardest ones, the difference amounts to double digits. This is definitely an impressive performance from the experimental multimodal extension, but it is not the model that is winning against Opus 4.8; rather, it is the one that is getting closer to it while losing in the majority of the comparisons. Also read: How Samsung redesigned the Galaxy Z Fold 8 Ultra and Fold 8 from the inside out: Sunghoon Moon explains Context is important here as well. Opus 4.8 is not some outdated model that DeepSeek is trying to take down. According to Anthropic's deprecation page, Opus 4.8 is marked as Active - fully supported and there are no plans for retiring it until May 2027. This makes the Opus 4.8 one of only five Opus-tier models that are in the same category. Therefore, DeepSeek chose a living example to compare with. The point is that nothing has been released regarding the comparison of the latest Opus generation and DeepSeek is comparing itself to something that is somewhat old already. Practically speaking, the new DeepSeek model supports JPEG, PNG, GIF and WebP formats and allows to upload up to 600 images within one request. It is quite an interesting upper limit for people who develop document and screenshot-based agents. DeepSeek also released 0.1.1 version of its agent harness with inbuilt support for the new model. Besides that, there is a Files API provided, which allows uploading an image once and referring to it multiple times using a file ID. The underlying theme here is always going to be cost, just as it always seems to be for DeepSeek. According to the reports, the cost of running this model is reported to be around 99% cheaper than that of Opus 4.8, which is the actual news that is hidden within the benchmark table. Any model that performs less in eight out of eleven tests but costs a fraction of the price would certainly become very interesting to developers. DeepSeek-V4-Flash-Vision-Exp is live on the company's API platform now. Whether "close to Opus 4.8'' really means anything in terms of reliable operations of an agent, specifically designed for this purpose, is the question that needs to be answered through testing of this benchmark table."