Source | Bohu Finance (bohuFN)

Recently, DeepSeek has made headlines with its significant price hikes, a move that is far from subtle.

Effective August 17, the DeepSeek-V4 series has introduced peak-valley pricing. Take V4-Pro as an example: the cost for cache-hit input has surged from 0.025 yuan per million tokens to 0.3 yuan during peak hours, marking a 12-fold increase or a staggering 1,100% hike. Similarly, cache-miss input has jumped from 3 yuan to 9 yuan, while output costs have soared from 6 yuan to 27 yuan.

Although a 50% discount is available during off-peak hours, making the scheme potentially more economical for batch processing and data cleaning tasks that can be scheduled during late-night hours, it poses a significant challenge for developers who require continuous debugging and repeated tool calls during regular work hours—precisely when peak pricing applies.

While the "maximum 1,100% increase" pertains to specific billing items rather than an overall 11-fold price hike, it still signals a paradigm shift: the DeepSeek myth, built on the foundation of ultra-low token pricing, is now being dismantled from within by escalating computing power costs.

Figure 1 | Price adjustments for three billing items in DeepSeek-V4-Pro; each panel employs an independent linear scale (Source: DeepSeek's public pricing, Reuters; chart by author)

Over the past year, DeepSeek has been grappling with two major "illusions."

The first illusion is that of affordability.

On paper, DeepSeek remains competitively priced. Even after the price hike, its absolute unit price is still lower than most overseas cutting-edge models. However, developers don't purchase tokens in isolation; they invest in a task that must be completed on time. When the pricing metric shifts from "price per token" to "total cost of completing a complex task," the conclusion changes dramatically.

In previous interviews with frontline developers, the author discovered that many perceive domestic models as "expensive" in real coding scenarios—not due to the specific token price, but rather due to the lack of cost-effective "packages" from model providers tailored for high-intensity coding.

High-intensity coding is not a one-time Q&A session; it involves the model continuously reading codebases, generating patches, running tests, retrying on failure, and consuming tokens at a rapid pace—with tens or even hundreds of billions of tokens consumed daily being a common occurrence.

OpenAI and Anthropic address this challenge with subscription packages. Codex is included in ChatGPT subscriptions, with Plus plans starting at $20/month and Pro plans at $100/month, offering higher usage quotas. Similarly, Claude Code is included in Claude Pro and Max plans, priced at $20, $100, and $200/month.

These plans are not truly unlimited—credits must be purchased or pay-as-you-go rates apply after reaching certain caps—but the fixed monthly fees cover a high-frequency usage range, encouraging developers to integrate models into their daily workflows without the need to calculate costs for every test run.

Figure 2 | Comparison of mainstream coding packages and DeepSeek's billing model (Source: OpenAI, Anthropic, Kimi, Zhipu official; chart by author)

DeepSeek, however, lacks a high-intensity coding package comparable to Codex or Claude Code, relying instead on pay-as-you-go APIs for core development scenarios. While peak-valley pricing improves computing power utilization, it does little to alter the decision-making logic of heavy users: the model is most needed when prices are highest.

Thus, under high-intensity coding conditions, pay-as-you-go DeepSeek feels "expensive" and "not cost-effective" to developers.

The second illusion is that DeepSeek has already secured the developer ecosystem.

Publicly, DeepSeek is one of the most recognized names in domestic large models; yet among developers surveyed for this article, it is rarely the first choice for critical programming projects.

According to discussions with developers, GPT and Claude are preferred for real coding scenarios, followed by domestic models like K3 and GLM.

Within the domestic model landscape, developers more commonly rank K3, GLM, and Qwen ahead of DeepSeek. DeepSeek-V4 Flash and Pro are more often utilized for writing, summarization, casual chat, or lightweight tasks with high error tolerance.

This is not to say that DeepSeek cannot write code—the evaluation criteria for coding have evolved.

A single code completion test assesses local generation capabilities; a multi-hour Agent task evaluates context retention, tool call stability, failure recovery, post-modification testing, and the ability to know when to stop. Benchmarks measure snapshot capabilities, but production environments evaluate end-to-end success rates. A few points' lead on a benchmark generates buzz; one fewer failure in production drives renewals.

Thus, DeepSeek's real challenge is not whether it remains "affordable" after the price hike but its lack of a production-grade position in the highest-willingness-to-pay, highest-stickiness coding scenarios. Low prices can attract users but cannot complete tasks for the model.

Facing cost pressures, DeepSeek has not continued to compete head-on.

In August, DeepSeek Harness entered developer preview. The official formula is straightforward: Agent = Model + Harness. The model handles reasoning; Harness places the model in real-world environments, enabling tool access, memory management, sandboxing, session management, scheduling, and sustained execution.

Its most radical feature is "everything is a plugin." Model adapters, tools, Skills, sessions, storage, Agent loops, scheduling, and even the UI can be replaced and recombined. The underlying Cordis mounts these capabilities in a shared runtime without requiring modifications to an untouchable core. Meanwhile, the system logs prompts, reasoning, tool calls, execution results, sub-Agent scheduling, and context injections in traceable, append-only logs, supporting recovery, forking, searching, and replay.

Figure 3 | Schematic of DeepSeek Harness's "everything is a plugin" architecture (Source: DeepSeek Harness official documentation; chart by author)

This is a clever strategic move.

If DeepSeek cannot outperform GPT or Claude in coding model capabilities anytime soon—or even consistently beat K3, GLM, and Qwen—it can aim for a higher-level position: running all models within its orchestration framework. As models upgrade, Harness remains on developers' machines; whoever controls Agent tool interfaces, execution logs, and plugin standards may become the next-stage AI ecosystem infrastructure.

This is why some media outlets have dubbed it the "Android of the AI era."

However, DeepSeek Harness is currently more akin to Linux or Docker than Android. Android's commercial value stems not just from open-source code but from closed-source GMS, Play Store, and default distribution rights.

DeepSeek Harness uses the MIT license and emphasizes model neutrality, lacking both closed-source services exclusive to DeepSeek and a plugin store or irreplaceable cloud hosting gateway. Developers can use it to call Claude, GPT, Kimi, or GLM—and the more successful the framework, the easier it may become to replace DeepSeek's own models.

Earlier this year, OpenClaw demonstrated this tension.

It rapidly built a massive open-source ecosystem, with cloud providers, model companies, and security firms launching products around it. Yet the project itself eventually moved to a foundation, and founder Peter Steinberger joined OpenAI. OpenClaw has not independently commercialized.

Open-source infrastructure can transform industries but does not naturally capture industry profits. Those controlling computing power, models, enterprise clients, or distribution channels are better positioned to convert ecosystem heat into cash flow.

Thus, Harness alone won't "win back" the coding market for DeepSeek anytime soon.

But viewing Harness as merely a simple Agent product underestimates its strategic value.

Its true worth to investors lies not in how many licenses it sells today but in its potential to become DeepSeek's gateway to the next technical paradigm: turning every real-world failure into material for the next iteration to read, compare, and validate.

This is precisely the next direction OpenAI, Anthropic, and domestic AI firms are pursuing: RSI (Recursive Self-Improving AI).

To understand Harness's real value, we must first ask: Why does RSI matter?

For years, the industry has improved models by stacking computing power, parameters, and data. But training freezes capabilities; faced with out-of-distribution tasks, it doesn't turn failures into inheritable experience—only waiting for the next data and manual training cycle. As data depletes and computing power grows costlier, this approach's marginal returns keep declining.

RSI is being bet on by giants because it aims to let AI identify problems, modify itself, and validate results in real tasks, closing the generalization gap.

At this stage, a more practical RSI is a gated engineering loop: record execution trajectories, identify stable weaknesses, modify prompts, tools, memory, or scheduling policies; candidate versions enter a sandbox, promoting only after passing independent evaluation and holdout set regression, or rolling back immediately otherwise.

Figure 4 | Closed loop from execution trajectories to verifiable Harness improvements (Source: *Self-Harness*, OpenAI, Anthropic; chart by author)

Harness sits at a critical loop point. Weights are expensive, uninterpretable black boxes; Harness is an external, readable, modifiable, rollbackable software layer. It defines how models call tools, save memory, and accept feedback—and is where complete trajectories are generated. Having AI modify its own execution protocol is far more practical than retraining its "brain."

Industry practice provides evidence. Anthropic incorporates full trajectories and environmental states into Agent evaluation; OpenAI engineer Ryan Lopopolo argues that baking tools, testing, and feedback loops into Harness creates compounding capabilities. Papers like *Self-Harness* find that modifying only Harness with fixed weights improves all nine experiments, with a maximum relative gain of 132%; Darwin Gödel Machine lets Coding Agents modify their own code and workflows, boosting SWE-bench scores from 20% to 50%. Capability growth can happen outside parameters.

DeepSeek Harness already features pluginization, append-only trajectories, modifiable Agent loops, and sandboxing. This makes it an "RSI-friendly foundation," providing an engineering basis for future Agent self-improvement.

Such industry-transforming changes are precisely the "strategic option" that capital is betting on in DeepSeek's financing.

According to Reuters, DeepSeek secured approximately 50 billion yuan in financing in June, planned another round at a valuation of around 500 billion yuan in July, and discussed listing on the STAR Market. Meanwhile, DeepSeek began recruiting IDC, energy, and civil engineering talent to expand its data center, Agent, and chip teams.

A senior large-model engineering veteran told the author that while algorithmic optimization will continue, the probability of replicating R1-scale cost reductions through algorithms alone has significantly declined under current technical paths. Engineering teams now see more certainty in increasing chips, memory, bandwidth, and cluster scale.

On the other hand, DeepSeek's longtime financial backer, High-Flyer Quantitative, isn't an infinite ATM: according to Simuwang data cited by media, as of late July, eight of the nine High-Flyer products on display had year-to-date negative returns.

Thus, DeepSeek's AGI journey can no longer rely on "low prices," "cost reductions," or High-Flyer's funding.

This is why Harness—a newer, sharper "spear"—has become DeepSeek's latest strategic option.

Figure 5 | Realization paths and value leakage points for Harness's strategic option; chart by author

But an option isn't a realization. If Harness remains a neutral MIT project, revenue and clients will still flow to model and cloud providers; DeepSeek must develop its own "GMS"—cloud hosting, enterprise evaluation, proprietary data pipelines, or a model-Harness synergy layer. If it can turn real failures into verifiable, inheritable capabilities for the next version, Harness could rewrite AI improvement; otherwise, it's just another prestigious but low-margin open-source framework.

R1 has redefined the cost of reasoning. What Harness needs to prove is whether DeepSeek can still define the next round of evolution.

The copyright of the cover image and accompanying pictures in the article belongs to the respective copyright owners. If the copyright holders believe that their works are not suitable for public viewing or should not be used free of charge, please contact us promptly, and this platform will make immediate corrections.