Feedback: Web search quality and token cost make DSH unusable for professional research

[elbarto1917]

I am writing as a professional user who was eager to adopt DeepSeek Harness (DSH) as a paid alternative to other AI subscription services. I believe in your model — it outperforms competitors I was paying $40/month for — and I wanted to redirect that budget to DeepSeek.

However, after extensive daily use of DSH, I must share feedback that I hope is taken constructively. The product currently falls short of DeepSeek's own free web version in critical areas, which makes it impossible to justify paying for it as a primary tool.

1. Web Search: Inferior to the Free Web Version

This is the most severe issue. DSH's web search is not just slower or less parallel — it is qualitatively worse. It fails to find what the free web version finds. It does not enter documents, does not cite sources with precision, and often requires repeating searches. I have spent money and time building custom search plugins with multiple meta-search engines to try to close this gap, i've also tried many community plugnis. They did not work. The result was slower (up to 15x), more expensive (60K–250K tokens per search), and still far below the free web version in quality.

By contrast, the free web version can search 40–80 pages in seconds, read them, and cite exact passages. That is the baseline I expect from a paid product.

2. Token Consumption: Unsustainable Cost per Response

This is not just "context accumulation" over a long conversation. A single response in DSH can exceed 1 million tokens. A moderately complex conversation can easily reach 100 million tokens. I spent $10 in two days using DSH for basic tasks, when my reference point was $40/month on a competitor. The cost-per-response makes DSH economically irrational even for tasks where it might technically perform well.

3. Product Incoherence: Paid Version Inferior to Free Version

It is illogical that a paid, specialized agent harness is worse than the free web product in search, cost, and speed. I am willing to pay more for a product that improves on every dimension. What I cannot accept is paying for something that is inferior in the basics. A flexible product would allow switching between a "web mode" (with full search) and an "agent mode" (for offline/document-heavy work) within the same project. DSH does not offer this.

4. Attempts to Fix It Have Failed

Community plugins (ModSearch, etc.) and my own custom developments have not solved the problem. They either compromise quality, increase time, or increase cost. The architectural gap between DSH and the web version appears too large to bridge with plugins alone.

5. Suggestions

• Improve the native web search to match or exceed the free web version's depth, speed, and citation accuracy.
• Implement efficient context management: aggressive compression, selective pruning, and strict per-response token budgets.
• Allow users to switch between "web mode" and "agent mode" within the same session or project.
• Provide transparent token consumption metrics per response so users can understand and control costs.
• Consider offering a paid tier that includes the full web search capability as a baseline feature, not an add-on.

I am not writing to complain. I am writing because I want to pay for this product. The model is excellent. The harness is not ready, and the cost structure makes it unusable for professional research work. I hope this feedback reaches the product team.

Thank you for your time and for the work you do.

[nomuteinthebook]

hi, for the advice one, if it means that the efficiency of dsh is Inferior to the Free Web Version of deepseek model itself ?

[elbarto1917]

Good question — I should clarify, because it's an important distinction.

No, the model itself is not inferior. The model is excellent, and that's exactly why I wanted to pay for it. The problem is the harness architecture around it, not the model's intelligence.

To be precise:

1. Web search is not limited to 4 queries because of a hard cap — that limit is modifiable. I built my own plugin to raise it to 40 parallel searches. It improved things slightly, but the result was still far below the free web version: hundreds of thousands of tokens per search, minutes of latency, and citations that were still less precise. Raising the limit does not close the gap, because the gap is in content extraction and document reading, not in the number of parallel calls. The web version, as I've said, can perform 30 searches and answer with precise quotes from scholarly documents in 3 to 30 seconds.

2. Token consumption seems to be a harness design issue, not a model issue. DSH re-sends the full conversation context on every step of the agent loop. A single response can exceed 1M tokens; a moderately complex conversation can reach 100M. I understand that full context can be useful in some scenarios, and I'm not saying it should be removed. But it should be natively adjustable: the user should be able to choose between full context, compressed context, or selective pruning. Right now the only options are unofficial plugins that either limit immediate context too aggressively or destroy previous context while synthesizing. Neither works well.

3. I want to be clear: Harness's reasoning is excellent. It has helped me enormously with very complex research and produced excellent internal results. The problem is the cost of reaching those results: hundreds of millions of tokens in a short time, enormous responses, and search quality so poor that I have to supply the sources myself or correct it constantly — at the cost of even more millions of tokens.

So when I say "efficiency," I mean end-to-end task efficiency:

• Quality of search results and citations.
• Time to complete a research task.
• Token cost per useful answer.

In all three, DSH is currently worse than the free web version, even though the underlying model is the same or better. The gap is in the harness: search implementation, context management, and tool orchestration.

That's why I framed it as a product coherence problem. A paid, specialized harness should not be worse than the free product in the basics.

I'm not criticizing the model — quite the opposite. It has notably surpassed paid models that were, until recently, considered superior, and that is precisely why I wanted to bring my subscription budget here. I genuinely believe this gap should be resolved for DSH to become a complete alternative for professionals in research-intensive fields like mine.

Thank you for your time and for the work you do.