DeepSeek V4 Pro Officially Released: Agent Benchmarks, API Pricing, and the New Coding Focus
DeepSeek has officially pushed its V4 Pro model from preview into a newer production version, giving developers access to the updated DeepSeek-V4-Pro-0813 model through the API. Th
DeepSeek V4 Pro Officially Released: Agent Benchmarks, API Pricing, and the New Coding Focus
Introduction
DeepSeek has officially pushed its V4 Pro model from preview into a newer production version, giving developers access to the updated DeepSeek-V4-Pro-0813 model through the API.
The source article presents the release as a major step for DeepSeek's agent and coding capabilities. Early benchmark results reported in the article put V4 Pro close to the leading frontier models, particularly on tasks involving coding, terminal interaction, cybersecurity agents, and long-horizon tool use.
At the same time, early users noticed an unusual change in the model's reasoning output. The reasoning text can look highly compressed and difficult for humans to read, suggesting that the model is increasingly optimized around machine-efficient reasoning rather than polished natural-language explanation.
The official DeepSeek documentation confirms that the current deepseek-v4-pro API identifier now serves DeepSeek-V4-Pro-0813, with the calling method unchanged.
DeepSeek V4 Pro Is Finally Here
The source article opens with a simple message: DeepSeek V4 Pro has officially arrived.
Early benchmark results highlighted in the article include HLE, Terminal-Bench, CyberGym, and DeepSWE. According to the source, V4 Pro was already outperforming Opus 4.8 on several of these agent-oriented evaluations and was approaching, or in some cases exceeding, the reported results of Fable 5.
The article attributes much of this improvement to post-training, arguing that the new release is not simply a larger base model but a model that has been heavily optimized for the kinds of tasks modern AI agents actually perform.
The broader message is clear: DeepSeek is no longer positioning V4 Pro purely around low cost. The Pro model is being aimed directly at the upper end of the model market.
DeepSeek V4 Pro-0813 Is Already Available Through the API
At the time of the original article's publication, DeepSeek had not yet published a corresponding official blog post, but the API platform had already been updated.
Developers could use the V4 Pro API and access the updated model identified as:
DeepSeek-V4-Pro-0813
DeepSeek's current official documentation confirms that the public model identifier remains:
deepseek-v4-pro
The service now routes that identifier to DeepSeek-V4-Pro-0813, so developers do not need to change the model name in their existing API integration.
The official API documentation also confirms support for OpenAI-compatible and Anthropic-compatible API formats. V4 Pro supports a 1M-token context window, tool calls, JSON output, and both thinking and non-thinking modes.
The Pricing Is Still Aggressive
The source article reports the following DeepSeek V4 Pro pricing:
| Item | Price per 1M tokens |
|---|---|
| Input | ¥3 |
| Cached input | ¥0.025 |
| Output | ¥6 |
These figures match DeepSeek's current official Chinese pricing documentation for deepseek-v4-pro. The official page lists ¥0.025 per million cached input tokens, ¥3 per million uncached input tokens, and ¥6 per million output tokens.
That pricing remains one of DeepSeek's biggest competitive advantages.
The interesting change is that DeepSeek is now trying to combine that price advantage with stronger agent performance. Earlier versions made the cost argument easy to understand; V4 Pro is attempting to make the performance argument just as compelling.
The Good News: Coding and Agent Performance Have Improved Sharply
The strongest part of the V4 Pro release, according to the source article, is coding.
The highlighted benchmark results cover tasks that are much closer to real software-engineering workflows than a simple question-and-answer benchmark.
HLE
The source article highlights Humanity's Last Exam as one of the evaluations where V4 Pro competes with leading models.
HLE is intended to test difficult knowledge and reasoning capabilities. The article uses the result as evidence that V4 Pro's improvements are not limited to straightforward code completion.
Terminal-Bench
Terminal-Bench evaluates models on tasks performed through a terminal environment.
This is particularly relevant for coding agents because the model must interact with a real tool environment rather than simply return a code snippet.
The source article reports V4 Pro as being close to the top of the benchmark and ahead of Opus 4.8.
CyberGym
CyberGym is another agent-oriented evaluation highlighted in the article.
The source describes V4 Pro as showing particularly strong performance here, suggesting that the model is becoming more capable at multi-step technical tasks involving tools and environments.
DeepSWE
DeepSWE is perhaps the most directly relevant result for developers.
The source article says V4 Pro significantly improves over the previous version on this evaluation and reaches a level close to the leading frontier models.
Taken together, these results support the source article's central argument: DeepSeek V4 Pro has been optimized heavily for coding agents and tool-based workflows.
But There Is a Strange Change in the Chain of Thought
The source article also reports a less positive observation from early users.
After testing V4 Pro, some users noticed that its chain-of-thought text had become extremely compressed and difficult to read.
One example reproduced in the source article contains fragmented, machine-like wording rather than normal human language.
The model can still convey useful information through the reasoning output, but the text is no longer necessarily pleasant or intuitive for a person to read.
This behavior is consistent with the broader direction of reasoning-model development.
DeepSeek's official documentation currently describes a dedicated thinking mode and exposes the reasoning output through reasoning_content in the API. It also provides controls for thinking mode and reasoning effort.
The source article speculates that less human-readable reasoning may help reduce token consumption while preserving useful information for the model itself. That interpretation should be treated as an observation rather than an official DeepSeek explanation.
The Trade-Off: Better Agents, Less Human-Friendly Reasoning
There is a natural trade-off here.
If a model spends its reasoning budget producing polished sentences, the intermediate process may be easier for humans to follow but less token-efficient.
If the model is optimized to solve a task as efficiently as possible, the internal reasoning representation may become much less readable.
For an AI coding agent, the second approach can be perfectly reasonable.
The agent does not necessarily need to write an essay while deciding which command to execute. It needs to reach the right action efficiently, inspect the result, and continue.
That is why the source article connects the change in reasoning style with the broader rise of agentic coding.
The model is increasingly optimized for doing, not necessarily for explaining every intermediate step in natural language.
What This Means for Writing and General Content Tasks
The downside is that optimization for coding and agents may not translate directly into better long-form writing.
The source article argues that V4 Pro's reasoning style may be less attractive for writing-heavy workflows, especially when the user expects polished language and transparent explanations.
This does not mean the model cannot write.
Rather, it reflects a change in priority. If most of the model's post-training effort is focused on coding, tool use, and agent reliability, some other capabilities may not improve at the same rate.
For developers, that distinction matters.
A model can be excellent for repository-level coding and still not be the best choice for every content-generation task.
DeepSeek's Bigger Shift: From Cost Leader to Frontier Competitor
The article places V4 Pro in the context of several recent releases from China's open-model ecosystem.
The source describes a rapid sequence of new models that have increasingly targeted the performance of leading closed models.
It mentions GLM 5.2, Kimi K3, and other recent releases as part of the same broader trend.
DeepSeek's earlier V4 Flash release focused heavily on cost efficiency. The source article describes that strategy as helping push API pricing downward across the market.
V4 Pro takes a different position.
The target is no longer simply:
"Can we provide similar capability for less money?"
It is closer to:
"Can we compete for the top performance tier while keeping the cost advantage?"
That is a much more ambitious position.
V4 Pro Is Also Built for Agent Integrations
DeepSeek's official documentation says V4 models are designed to work with popular AI agent and coding-assistant tools, including Claude Code, GitHub Copilot, and OpenCode. Developers can use DeepSeek as the backend model for these tools.
This is important because model capability is increasingly determined by how well a model works inside an agent loop.
A strong coding model needs more than code-generation ability. It needs to:
- Understand a large codebase.
- Plan a multi-step change.
- Call tools reliably.
- Inspect command output.
- Recover from errors.
- Continue across multiple turns.
- Produce a final implementation that actually works.
The V4 Pro release is clearly aimed at this workflow.
One More Thing
The source article closes on a more emotional note.
It compares the release with the DeepSeek moment of early 2025 and argues that the model's positioning has changed.
Earlier, DeepSeek's most obvious weapon was price-performance.
Now the company is trying to put V4 Pro directly into the conversation around the strongest available models.
That makes the release important even beyond the model itself.
If an open model can reach the frontier on agentic coding while maintaining comparatively low API prices, the pressure on closed-model providers becomes much stronger.
常见问题
What is DeepSeek V4 Pro-0813?
DeepSeek-V4-Pro-0813 is the current updated version served through the deepseek-v4-pro model identifier. DeepSeek's official API documentation confirms that developers can continue using deepseek-v4-pro without changing their integration.
How much does DeepSeek V4 Pro cost?
DeepSeek's current Chinese pricing page lists ¥3 per million uncached input tokens, ¥0.025 per million cached input tokens, and ¥6 per million output tokens for V4 Pro. Prices can change, so the official pricing page should be checked before deployment.
What is DeepSeek V4 Pro mainly optimized for?
The release places strong emphasis on coding, agentic workflows, tool use, and reasoning. DeepSeek's official documentation also lists tool calls, 1M context, thinking mode, and integrations with popular coding-agent tools.
Does DeepSeek V4 Pro have a 1M-token context window?
Yes. DeepSeek's current official model documentation lists a 1M context length for V4 Pro. It also lists a maximum output length of 384K tokens.
Why does DeepSeek V4 Pro's reasoning look unusual?
The source article reports that early users saw highly compressed, machine-like reasoning text. The article suggests this may be related to efficiency, but DeepSeek's official documentation does not state that the unusual style was specifically introduced for that reason.
Can DeepSeek V4 Pro be used with coding agents?
Yes. DeepSeek officially documents integrations with tools such as Claude Code, GitHub Copilot, and OpenCode. It also supports OpenAI-compatible and Anthropic-compatible API formats.
Is DeepSeek V4 Pro open source?
DeepSeek's V4 preview release was announced as open-sourced, including open weights and a technical report. The current API service and model version are documented separately from the release materials, so developers should distinguish between open model artifacts and the hosted API service.
相关工具
- DeepSeek API: Official API endpoint for accessing DeepSeek models.
- DeepSeek API Documentation: Official documentation for model access, pricing, tools, and integrations.
- Claude Code: An AI coding agent that can use DeepSeek as a backend through supported integrations.
- OpenCode: An open-source coding agent supported by DeepSeek's documented agent integrations.
- GitHub Copilot: An AI coding assistant listed by DeepSeek among supported agent integrations.
Related Links
- DeepSeek Models & Pricing: Official current model specifications and API pricing.
- DeepSeek First API Call: Official guide for calling
deepseek-v4-proanddeepseek-v4-flash. - DeepSeek Thinking Mode: Official documentation for enabling and controlling reasoning mode.
- DeepSeek Agent Integrations: Official documentation for using DeepSeek with coding and agent tools.
- DeepSeek V4 Preview Release: Official V4 release announcement and technical overview.
- DeepSeek GitHub: Official DeepSeek GitHub organization.
Summary
DeepSeek V4 Pro marks a shift in the company's strategy. The new Pro model is not being positioned only as a low-cost alternative; its main pitch is increasingly about competing at the frontier of coding and agentic performance.
The current API version, deepseek-v4-pro, serves DeepSeek-V4-Pro-0813, with 1M context, tool calls, thinking modes, and aggressive pricing.
The reported changes in reasoning style also show where the model's priorities are moving: less emphasis on human-readable intermediate text, and more emphasis on efficient reasoning and agent execution.