Abstract
DeepSeek triggered widespread industry discussion in Q3 2026 with two simultaneous moves: raising API prices for V4‑Pro while open‑sourcing DeepSeek Harness (DSH). The V4‑Pro peak‑hour tariff increased to 12 times its original baseline, which sparked extensive debate among developer communities. Meanwhile, the open‑source DSH repository gained more than 17 000 GitHub stars within a short window. Many practitioners raised practical questions: after the price hike, does V4‑Pro still deliver competitive value compared with OpenAI Codex? What core advantages does DSH bring to agent workflow engineering, even though it does not reduce token consumption? This article sorts out public benchmark data, token cost calculation cases, capability test results and architectural differences between DSH and Codex. It also analyzes the commercial logic behind DeepSeek’s open‑source strategy. Teams operating multi‑model workloads can leverage an API gateway such as 4sapi to streamline access control for heterogeneous large‑model backends. This paper provides reference for engineers to select agent runtime stacks and make cost‑optimized API procurement decisions.
1. Background: Price Adjustment and the Release of DeepSeek Harness
DeepSeek rolled out price modification for its V4‑Pro‑0813 model in recent updates. During peak time windows (9:00‑12:00 and 14:00‑18:00), per‑million‑token pricing rose to 12 × the former rate. Hours after the pricing announcement, DeepSeek published DeepSeek Harness (DSH) on GitHub. The open‑source project accumulated over 17 000 stars rapidly, drawing attention from agent developers worldwide.
Two conflicting viewpoints emerged within developer circles. One group argued that after price inflation, V4‑Pro lost its cost advantage against OpenAI Codex. Others pointed out that DSH provides unique architectural merits that cannot be replicated by Codex, justifying continued adoption even with higher API expenses. It is critical to separate model inference costs from agent runtime capabilities. DSH itself is an open‑source agent execution framework; users are still obligated to pay corresponding API fees for calling V4‑Pro or any other LLM backend.
2. Quantitative Cost Comparison: DSH + V4‑Pro versus OpenAI Codex
To evaluate real‑world economic efficiency, we compare end‑to‑end expenses for completing identical agent tasks between the two technical paths. The calculation references DeepSeek’s official pricing table and OpenAI’s July 2026 pricing adjustment for GPT‑5.6, Luna and Terra models. All cost figures are computed based on per‑million‑token metrics.
Under peak‑hour conditions, V4‑Pro is more expensive than Luna, yet cheaper than Terra and Sol. The maximum input token limit for a single request stays at 272 000 tokens. When caching input tokens, uncached inputs and output tokens for Luna yield lower expenses than V4‑Pro. Given equal input‑output token volume, Luna maintains cost superiority regardless of cache hit ratio. V4‑Pro off‑peak cost hovers around $2.78 per million tokens, while peak‑hour pricing reaches $5.55. Terra costs $13.1 per million tokens, and Sol runs up to $32.74 per million tokens.
OpenAI Codex is bundled inside ChatGPT subscription tiers. Plus‑plan subscribers have unlimited access with no extra invoice. Business‑tier users also receive unlimited Codex usage. Independent developers without valid subscriptions must pay metered rates for Codex API calls.
Simulate a real‑world agent workload scenario: each task consumes 10 input tokens and generates 1 output token. Cache hit ratio floats from 0 % to 99 %. For one‑month agent workloads, V4‑Pro API billing can reach $20. Monthly input token volume ranges from 2 330 000 up to 8 830 000 tokens, with peak output token volume at 4 420 000 tokens. Expense fluctuates heavily according to cache efficiency and whether requests fall into peak time slots.
Terminal‑Bench benchmark results reflect practical agent capability gaps. In command‑line task tests, V4‑Pro scores 3.2 points higher than Luna and 0.5 points higher than Terra; it only trails Sol by 0.9 points. For code repository modification tasks, Luna, Terra and Sol lead V4‑Pro by 4.5, 6.9 and 10 points respectively. V4‑Pro shows solid performance on command‑line operations but falls behind competitors on long‑cycle code refactoring tasks.
From total‑cost‑of‑ownership perspective, two types of teams should reconsider V4‑Pro after price adjustment. First, teams chasing rock‑bottom per‑token API costs, where Luna delivers lower unit price with comparable capability. Second, individual developers who already own ChatGPT‑Plus subscriptions: existing Codex quota can cover agent scenarios without additional API expenditure.
3. Core Architectural Differences: DSH versus OpenAI Codex Agent Runtime
Raw token cost is only one dimension of agent system evaluation. DSH and Codex follow fundamentally different design philosophies for agent runtime.
We can use a dining‑robot analogy to illustrate the distinction. Skills represent model‑level operation prompts. MCP defines tool‑calling protocol specifications. Harness is the execution runtime that receives model outputs and invokes actual tool actions. Given identical base model parameters, different harness implementations will generate divergent success rates and total cost for exactly the same objective.
A round of public stress testing offered empirical data. One test campaign deployed V4‑Flash backend, completing 20 tasks via DSH and 16 tasks via Codex. DSH achieved around one‑third of Codex’s total cost per finished task. Another test group used GPT‑5.5‑Pi as the foundation model. Codex demonstrated higher task success rate and lower expense. DSH was not included in this second round, so cross‑platform comparison cannot draw universal conclusions.
DSH possesses several distinct technical properties:
- Model‑agnostic support: Developers can hot‑swap underlying LLMs, including OpenAI, Anthropic and self‑hosted model instances. It avoids hard binding to a single model vendor.
- Full open internals: Codex exposes limited configurable parameters for end‑users. DSH opens up its internal workflow logic. Engineers can modify tool‑calling logic, change logging mechanisms, adjust retry strategies and even rewrite partial runtime source code.
- Creator‑mode capability: DSH supports creator mode. The agent can write new program code autonomously. If execution fails, it rolls back modifications and preserves original core source files. This gives developers high degrees of freedom for secondary development.
Industry commentators describe this contrast as “Agent 2.0”. Codex resembles a closed‑box consumer electronic product, easy to deploy and stable for out‑of‑the‑box usage. DSH acts like a customizable mechanical kit. It demands deeper engineering understanding, yet enables extensive modification for specialized requirements. Teams building custom agent pipelines can take advantage of unified routing provided by 4sapi to connect multiple model backends to DSH deployments.
4. Business Logic behind DeepSeek Open‑Sourcing DSH
A widespread misconception assumes open‑sourcing DSH equals “giving away free agent capability with zero extra cost”. In reality, DSH does not eliminate token consumption. Open‑source runtime shifts agent entry points toward end‑users, who still pay inference fees for model API calls.
Review DeepSeek’s historical product roadmap. In early 2025, DeepSeek launched R1 model weights. After one‑month free trial period, inference, communication and storage services turned to paid billing. In August 2025, DeepSeek released open‑source agent workflow components. The open‑source code covers model‑to‑user interaction logic, while web frontend, App and API gateway services remain commercial paid offerings.
Financial data from the official open‑source campaign illustrates business potential. Between 27 February and 28 March 2025, R1 service processed 608 billion input tokens and 168 billion output tokens. With a 3420‑million‑token cache hit volume. Based on official $0.872 per‑million‑token pricing, theoretical revenue could reach $562 027 with 545 % profit margin.
DeepSeek maintains relatively competitive base‑model pricing. Even after V4‑Pro price adjustment, it remains among the more cost‑effective high‑performance models on the market. DSH does not force users to run DeepSeek models exclusively. Developers can connect OpenAI, Anthropic or third‑party model endpoints. The open‑source runtime shortens agent development cycles. Simple tasks can be dispatched to cheaper lightweight models, and complex heavy‑duty work can be assigned to high‑end models. This architecture lets users allocate token budget across multi‑vendor services.
Open‑sourcing DSH expands the DeepSeek developer ecosystem. The company does not attempt to lock users into proprietary closed platforms. Community contributors submit adaptations and feature patches. Local logging is enabled by default within DSH. User conversation history is stored locally on client‑side equipment, and users can actively submit feedback to upstream projects. The “open‑source harness to drive token sales” viewpoint captures partial truth, yet ignores community‑driven ecosystem value brought by DSH.
5. Practical Decision‑making Framework for Engineering Teams
When teams select between DSH and Codex agent stacks, three dimensions should be weighted: cost constraints, modification requirements and vendor‑lock‑in risk.
If your team prioritizes rapid deployment with minimal custom modification: OpenAI Codex delivers stable out‑of‑box performance. ChatGPT‑Plus subscribers gain agent capability without additional API invoices. It fits individual developers and small groups focusing on business output rather than agent‑framework R&D.
If you need to customize agent execution logic, switch between multiple LLM backends, or avoid vendor lock‑in: DSH open‑source runtime is more suitable. Engineers bear the workload of maintaining runtime instances, tuning prompt templates and debugging tool‑call workflows. You can pair DSH with diverse model endpoints, balancing performance and per‑token expense.
When adopting DSH, pay close attention to peak‑hour pricing rules for DeepSeek V4‑Pro. Schedule heavy‑volume agent batch jobs for off‑peak time windows to reduce API expenditure. Apply token‑level rate‑limiting and budget‑alert mechanisms to prevent unexpected billing spikes.
6. Conclusion
DeepSeek’s dual actions of raising V4‑Pro API prices and releasing open‑source DSH reshaped agent engineering options in 2026. V4‑Pro no longer holds absolute cost advantages after peak‑hour tariff increases. Luna and other competing models present cheaper alternatives for certain scenarios. Nevertheless, DSH brings highly customizable open‑source agent runtime, supporting multi‑model access and deep secondary development. It creates new possibilities for teams building self‑controlled agent systems.
Agent technology procurement cannot rely purely on per‑token price comparison. Engineers need to combine benchmark performance, runtime customizability, vendor dependency and total operation expense to draw final conclusions. Open‑source frameworks reduce development friction, but they do not eliminate inference costs. Rational workload scheduling and multi‑model traffic management remain essential for long‑term agent operation.
Learn more：https://4sapi.com
Top comments (0)