The main focus is not simply a new model label. DeepSeek says the release substantially improves performance in production AI-agent workflows, including coding, tool use, and multi-step task execution.

What the V4 Pro API release changes

Existing API calls continue to work

The version update does not require developers to replace the model identifier in their applications. DeepSeek’s documentation says that setting the model to deepseek-v4-pro will use the latest generally available version.

That makes the release relatively straightforward for existing integrations: the model name stays stable while the underlying API endpoint begins serving the 0813 release.

Responses API and Codex support are built in

DeepSeek says the API now supports the OpenAI Responses API format and is optimized for Codex workflows, including one-click Codex configuration. The broader API documentation also lists compatibility with OpenAI Chat Completions and Anthropic interfaces.

For developers, the practical significance is compatibility with agent-oriented application patterns rather than a basic text-completion endpoint. Responses-style interactions and Codex integration are designed for applications that need to combine model output with tools, code execution, and repeated steps.

Reported benchmark performance

DeepSeek’s published results show a large improvement over the V4 Pro preview on several agent and coding evaluations. On Terminal-Bench 2.1, V4 Pro-0813 scored 87.9, compared with 72.1 for the preview version. That placed it just below Fable 5’s reported 88.0.

Other reported results include:

• CyberGym: 83.3.
• DeepSWE: 62.7, up from 12.8 for the preview version.
• NL2Repo: 61.5.
• AutomationBench: 31.8 in the published results cited in the launch coverage.

Coverage of the release says V4 Pro exceeded Fable 5 on CyberGym and AutomationBench while remaining narrowly behind it on Terminal-Bench. Those comparisons should be read carefully: the figures are principally based on DeepSeek’s own evaluation table and accompanying launch reporting, not a single independently administered benchmark study.

The results nevertheless show where DeepSeek is directing its effort: software engineering agents, terminal tasks, cybersecurity-related work, and workflows that require a model to complete several actions rather than produce one isolated answer.

V4 Pro API pricing

The supplied pricing reference lists the following rates per 1 million tokens:

• Cached input: $0.022 during off-peak periods and $0.044 during peak periods.
• Uncached input: $0.66 off-peak and $1.32 peak.
• Output: $1.98 off-peak and $3.96 peak.

The same reference says off-peak pricing is half the peak rate, with peak windows from 01:00–04:00 UTC and 06:00–10:00 UTC.

Pricing information in the supplied sources is not fully consistent across currency and documentation references: another report citing DeepSeek’s API documentation lists rates of 3 yuan per million input tokens, 6 yuan per million output tokens, and 0.025 yuan for cached-hit input. Developers should therefore confirm the currently displayed rates and applicable currency directly in DeepSeek’s API documentation before estimating production costs.

What this means for developers

The release is most relevant to teams building coding agents and tool-using systems. Existing users can keep the deepseek-v4-pro identifier, while Responses API and Codex compatibility may reduce the work required to connect the model to established agent tooling.

The benchmark results also make V4 Pro more competitive on the specific tasks that matter to autonomous software workflows. But a near-parity score on one evaluation does not establish that the model will perform identically across every coding, reasoning, or production workload. The reported results should be treated as DeepSeek’s claims until independently reproduced.

In short, DeepSeek’s V4 Pro announcement is an API and agent-platform upgrade as much as a model release: the model name remains stable, the integration surface is broader, and the reported agent scores have moved sharply higher from the preview version.

The main focus is not simply a new model label. DeepSeek says the release substantially improves performance in production AI-agent workflows, including coding, tool use, and multi-step task execution.

What the V4 Pro API release changes

Existing API calls continue to work

The version update does not require developers to replace the model identifier in their applications. DeepSeek’s documentation says that setting the model to deepseek-v4-pro will use the latest generally available version.

That makes the release relatively straightforward for existing integrations: the model name stays stable while the underlying API endpoint begins serving the 0813 release.

Responses API and Codex support are built in

DeepSeek says the API now supports the OpenAI Responses API format and is optimized for Codex workflows, including one-click Codex configuration. The broader API documentation also lists compatibility with OpenAI Chat Completions and Anthropic interfaces.

For developers, the practical significance is compatibility with agent-oriented application patterns rather than a basic text-completion endpoint. Responses-style interactions and Codex integration are designed for applications that need to combine model output with tools, code execution, and repeated steps.

Reported benchmark performance

DeepSeek’s published results show a large improvement over the V4 Pro preview on several agent and coding evaluations. On Terminal-Bench 2.1, V4 Pro-0813 scored 87.9, compared with 72.1 for the preview version. That placed it just below Fable 5’s reported 88.0.

Other reported results include:

• CyberGym: 83.3.
• DeepSWE: 62.7, up from 12.8 for the preview version.
• NL2Repo: 61.5.
• AutomationBench: 31.8 in the published results cited in the launch coverage.

Coverage of the release says V4 Pro exceeded Fable 5 on CyberGym and AutomationBench while remaining narrowly behind it on Terminal-Bench. Those comparisons should be read carefully: the figures are principally based on DeepSeek’s own evaluation table and accompanying launch reporting, not a single independently administered benchmark study.

The results nevertheless show where DeepSeek is directing its effort: software engineering agents, terminal tasks, cybersecurity-related work, and workflows that require a model to complete several actions rather than produce one isolated answer.

V4 Pro API pricing

The supplied pricing reference lists the following rates per 1 million tokens:

• Cached input: $0.022 during off-peak periods and $0.044 during peak periods.
• Uncached input: $0.66 off-peak and $1.32 peak.
• Output: $1.98 off-peak and $3.96 peak.

The same reference says off-peak pricing is half the peak rate, with peak windows from 01:00–04:00 UTC and 06:00–10:00 UTC.

Pricing information in the supplied sources is not fully consistent across currency and documentation references: another report citing DeepSeek’s API documentation lists rates of 3 yuan per million input tokens, 6 yuan per million output tokens, and 0.025 yuan for cached-hit input. Developers should therefore confirm the currently displayed rates and applicable currency directly in DeepSeek’s API documentation before estimating production costs.

What this means for developers

The release is most relevant to teams building coding agents and tool-using systems. Existing users can keep the deepseek-v4-pro identifier, while Responses API and Codex compatibility may reduce the work required to connect the model to established agent tooling.

The benchmark results also make V4 Pro more competitive on the specific tasks that matter to autonomous software workflows. But a near-parity score on one evaluation does not establish that the model will perform identically across every coding, reasoning, or production workload. The reported results should be treated as DeepSeek’s claims until independently reproduced.

In short, DeepSeek’s V4 Pro announcement is an API and agent-platform upgrade as much as a model release: the model name remains stable, the integration surface is broader, and the reported agent scores have moved sharply higher from the preview version.