DeepSeek introduces Harness: an agent framework with four operating modes

DeepSeek has made its V4 Pro model publicly available, releasing it under the name DeepSeek-V4-Pro-0813 on its pricing page. The flagship model is equipped with a context window of 1 million tokens and a maximum output of 384 thousand tokens, defaulting to a thinking mode, but also providing an endpoint without thinking mode for latency-sensitive requests.

The combination of a 1 million token input and 384 thousand token output allows for processing and generating a significantly larger volume of material within a single request compared to the previous generation, which is particularly relevant for working with large codebases, extensive document corpora, and multi-stage agents.

DeepSeek has also integrated features that most agent developers expect: structured JSON output, tool calling support, Responses API, and compatibility with Anthropic-API. Furthermore, beta support for conversation prefix continuation and FIM completion has been implemented, with FIM limited to the non-thinking mode.

At the service level, endpoints compatible with both OpenAI and Anthropic are offered, allowing teams already using either standard to switch without needing to rebuild their infrastructure.

The pricing reflects a deliberate separation of tiers. The Pro model costs 0.025 yuan per million tokens when cached, 3 yuan for cache misses on input, and 6 yuan on output. In contrast, the Flash model is priced at 0.02, 1, and 2 yuan respectively, making Pro approximately three times more expensive than Flash for input and output cache misses.

Differences are also highlighted by parallelism limitations: Flash supports 2500 concurrent requests, whereas Pro is limited to 500. This ratio indicates a clearer product structure for the V4 family. Flash is geared towards high-frequency, high-throughput traffic where cost dominates, while Pro concentrates power on heavier reasoning tasks, long-context coding, and agent workflows, where the model has a greater performance reserve for use.

Nevertheless, the absolute price of Pro remains moderate by the standards of advanced models, allowing DeepSeek to offset costs after the recent price increase announcement while remaining cheaper than most foreign competitors. Thus, the V4 lineup presents a two-pronged offering: a low-cost, high-throughput Flash tier for scaling, and a more powerful Pro tier aimed at coding and agent tasks.

The DeepSeek V4 Flash model demonstrated massive demand, consuming 8 trillion tokens in a single day on the OpenCode tool. This volume exceeded the average daily usage of the OpenRouter platform, which is around 6.6 trillion tokens. This surge in usage reflects a fundamental shift in AI interaction methods: where tasks used to be simple question-and-answer exchanges, agents now perform more complex operations such as reading files, modifying code, running programs, and retrying after errors, which significantly increases token consumption.

An example of such complex interaction was shown by a user who created a single-page 3D game project using Claude Opus 5. This process resulted in one HTML file that required 690 million tokens and cost nearly 3000 yuan.

With the advent of agent-driven tasks, the evaluation criterion for models has shifted from the quality of a single response to the cost of completing a long-running task. For instance, one million output tokens cost 2 yuan with DeepSeek, whereas the same operation with Claude Opus 4.8 cost $25 USD, equivalent to approximately 170 yuan, demonstrating an 85-fold price difference. Although the V4 Flash Max model scored 50 points compared to 56 for Opus Max in the Artificial Analysis Intelligence Index v4.1, the cost to pass this test was $72.02 for V4 Flash versus $3752.55 for Opus, with Opus gaining 6 extra points at roughly a 52-fold increase in call costs.

V4 Flash cannot yet prove absolute superiority, but if both models appear in the same candidate list, the 85-fold price difference could change the standard selection. When a model only slightly outperforms its competitor but costs dozens of times more, the mathematics of choice no longer favors the more expensive option.

The first to feel the impact of this trend are not the weakest models, as smaller models remain suitable for routing, and flagship models continue to handle the most complex tasks. Mid-tier models are the most vulnerable—they are stronger than V4 Flash but cost dozens of times more and are not powerful enough to solve the tasks that V4 Flash handles. This is DeepSeek's 'killer line': models do not necessarily need to be the strongest; they just need to be good enough to perform most tasks, leveraging a price gap of almost two orders of magnitude to displace more expensive options from the default list.

In terms of pricing, Opus costs $25 per million output tokens, Fable 5 costs $50, and Chinese models include Qwen 3.8 Max priced at 36 yuan, Kimi K3 at 100 yuan, and GLM 5.2 at 28 yuan.

The ability to sell Mutai at mineral water prices is due to the direct inference cost per token, which depends on activated parameters, accuracy, attention, KV cache, output length, hardware utilization, and parallelism. V4 Flash has 284 billion total parameters but activates about 13 billion per token by using hybrid attention, low-precision weights, and cache optimization. With a context of one million tokens, the inference computational load is about 10% of V3.2, and the KV cache size is about 7%. The V4 series overhauled the attention mechanism, implementing CSA compress-then-focus and HCA full-browse mechanisms, and utilizes DeepSeekMoE routing, directing each token to a small subgroup of experts. Commercial pricing and service costs follow these technical features. Thus, the 'killer line' redefines the boundary of development: the combination of performance and price becomes the decisive factor, where the cost of task execution is the determining metric.