DeepSeek V4 is Being Tested Again: Said to Be Approaching Opus 4.5 But Won't Be Cheap
With about a week left in the month, the biggest anticipation is for DeepSeek V4, which appears to be getting closer and closer, with the official side also making frequent moves. It has now been discovered that the API is also upgrading models, with good performance. Since announcing the launch of a test version on February 8th, DeepSeek V4 has actually been quietly changing, but the official side has not detailed the upgrade content each time.
Recently, the DeepSeek V4 page was upgraded with three modes available: fast, expert, and visual. However, previous upgrades were mainly for the web version, and now it has been discovered that the API-side model is also switching.
However, the specific name of this model is uncertain, and analysis suggests that the smaller parameter DeepSeek V4 Lite is the most likely.
Of course, there are also surprises. A big shot in the Linux.do community mentioned internal testing information, pointing out that DeepSeek will have 2 models, adjustable thinking gears, support FP8, and strong instruction following.
More importantly, it is mentioned here that its performance is close to Opus 4.5, but it is not as good as GPT in complex tasks.
Although Anthropic has released Opus 4.6 and 4.7, these two large models have been controversial recently, with issues such as reduced intelligence and increased usage directly triggering community backlash. If the larger model of DeepSeek V4 can reach the level of Opus 4.5, it will definitely be a leader among open-source large models. This capability is sufficient for the vast majority of developers, especially since it is a domestic open-source product, with far fewer restrictions than A-family.
However, everyone should also be prepared that as the performance of DeepSeek V4 improves, the cost will not be as low as before, and the price of the larger parameter model will not be cheap.
Although the specific price is not mentioned here, it will definitely be much higher than DeepSeek V3.X, but it should not reach the level of Opus, and the overall cost-performance ratio will still be there.