DeepSeek developed a competitor to Claude Code and subsequently raised its prices fourfold.

      Begin with the financial aspect, as it's the element that every developer has noticed. Starting from August 16, the price for one million output tokens via DeepSeek-V4-Pro has increased from $0.87 to a peak of $3.96. V4-Flash saw a price rise from $0.28 to $1.32, according to Bloomberg. Off-peak prices are set at half the new peak, resulting in costs of $1.98 for V4-Pro and $0.66 for V4-Flash.

      Comparing these discounted rates to last week's full price, off-peak V4-Pro now costs $1.98, significantly higher than the previous peak price of $0.87. The new discounted rate is over double the old undiscounted cost. Consequently, a team that transfers all workloads to off-peak times still pays more than twice the amount they would have during peak hours.

      Initially, the company focused on developing a competitor to Claude Code. The pricing gains additional clarity when considering what was introduced alongside it. On Thursday, DeepSeek unveiled a developer preview of DeepSeek Harness v0.1.

      A harness is a framework surrounding a model that allows an agent to perform actual tasks. It can read files, edit code, browse the internet, and continue until the job is complete. This is the layer marketed as Claude Code by Anthropic, where the revenue in agent-driven coding resides. The model acts as the engine, with the harness functioning like the car.

      DeepSeek had been hinting at this development for several days. It established a “DeepSeek Harness Team” account on WeChat and shared job postings for this team, reported by Bloomberg earlier in the week. One job description plainly stated the aim: the company intends to transform its models into state-of-the-art agent-driven products.

      This account is registered under a Beijing entity linked to DeepSeek, as confirmed by Tencent. For a company that primarily communicates through model releases, the establishment of a public team account effectively serves as an announcement.

      The choice of design is particularly noteworthy. DeepSeek claims its harness employs an open architecture, allowing users to incorporate any components they wish, including models from rival companies. This is contrasted with American competitors, which they say employ hard-coded products. Whether this characterization of Claude Code is accurate is debatable, but the underlying strategy is clear.

      An open harness that accommodates any model aims to dominate the workspace rather than the engine. If developers utilize your tool, the model underneath becomes interchangeable, even with your own.

      The industry has observed this layer evolve into a competitive field. Cursor has adjusted its prices in its own shift toward harness use, while Alibaba has internally banned Claude Code due to tracking concerns.

      Quietly, the flagship exited its preview phase, and then the notice disappeared. Beneath both narratives lies a third aspect. DeepSeek-V4-Pro-0813 was rolled out as the general-availability version this week, concluding a nearly four-month preview. The company made a brief announcement on its website claiming the model had "significantly enhanced agent capabilities." However, by Thursday afternoon, that statement was gone, according to the South China Morning Post. DeepSeek has not provided an explanation, nor did it respond to Bloomberg’s inquiry regarding the harness team.

      Typically, a company does not retract a statement about its flagship model in the same week it hikes the price. Developers reacted unfavorably, with early responses to the 0813 build suggesting disappointment regarding overall capabilities and dissatisfaction with the increased pricing, although researchers expressed admiration in specific areas like cybersecurity.

      The benchmark figures cited are provided by the vendor, and no independent evaluations have corroborated them for this latest build. DeepSeek’s own model card positions V4-Pro at an 80.6% score in its maximum reasoning setting for SWE-bench Verified, comparable to Gemini 3.1 Pro and slightly trailing Claude Opus 4.6 at 80.8%. In other assessments, however, it falls short. The card indicates a score of 67.9% on Terminal Bench 2.0 compared to GPT-5.4, which is at 75.1%, and 37.7% on Humanity’s Last Exam against Gemini 3.1 Pro at 44.4%.

      Thus, the price has increased more than four times for a model that its creator claims is roughly on par with leading standards for coding yet lags in other aspects. The underlying technology is undisputed; V4-Pro is a mixture-of-experts system with 1.6 trillion parameters, utilizing 49 billion active per token. DeepSeek asserts that its attention design reduces the compute needed for a single token to 27% of its predecessor’s requirements.

      This efficiency is precisely what allowed the previous prices to be feasible, which also implies that the substantial increase appears to be a decision based on profit margins rather than indicative of cost increases.

      When it comes to pricing,