# DeepSeek V4 Pro 0813

DeepSeek V4 Pro 0813 pins the checkpoint that took V4 Pro out of preview. It keeps the 1.0M tokens context window and hybrid attention of the series, with agentic behaviour that improved most in production settings.

- **Model ID:** `deepseek/deepseek-v4-pro-0813`
- **Type:** chat
- **Providers:** deepseek, baseten, gmicloud, runinfra, fireworks, alibaba, parasail, digitalocean, streamlake, deepinfra, novita
- **Context window:** 1,048,576
- **Maximum output tokens:** 1,048,576
- **Pricing:** $0.6/1M input tokens, $1.9/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/deepseek-v4-pro-0813

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'deepseek/deepseek-v4-pro-0813',
  prompt: 'Why is the sky blue?'
})
```

## About

DeepSeek V4 Pro 0813 was released August 12, 2026 as the general-availability checkpoint of DeepSeek's V4 Pro, ending a preview period that ran close to four months. DeepSeek describes it as superseding the preview build, with greatly enhanced agentic capabilities and improvements that show up most in production environments.

The architecture is a mixture-of-experts design with 1.6 trillion total parameters and 49 billion active per token, combining two attention variants DeepSeek calls Compressed Sparse Attention and Heavily Compressed Attention. Together they cut per-token inference compute and KV cache substantially at the million-token setting, which is what makes a context window of 1.0M tokens practical rather than theoretical.

DeepSeek V4 Pro 0813 keeps thinking and non-thinking modes, tool calls, and up to 1.0M tokens per response. On vendor-reported results it scores 87.9 on Terminal Bench 2.1, 74.1 on Toolathlon-Verified, and 80.6 on SWE-bench Verified. Independent evaluators have not replicated those figures for this build.

You can integrate DeepSeek V4 Pro 0813 through AI SDK, Chat Completions API, Responses API, Messages API, or other API formats, from TypeScript or Python.

## What to consider

Pricing on this model changed shortly after release, moving to a schedule with peak and off-peak rates where off-peak bills at half the peak rate. Cache-hit input saw the largest increase. Check the pricing panel on this page before sizing a workload, and if your traffic is shiftable, note that the same job can cost meaningfully less outside peak hours.

Independent reception has been more mixed than the vendor benchmarks suggest. DeepSeek V4 Pro 0813 scores 53 on the Artificial Analysis Intelligence Index, level with GLM-5.2 and behind the mid-tier of the GPT-5.6 series, and reviewers have been cooler on general capability than on its niche strengths. The headline evaluation numbers are vendor-reported and have not been independently replicated for this build, so benchmark it on your own tasks.

The dated slug pins a specific checkpoint. Use it when you need reproducible behaviour, and use the undated `deepseek-v4-pro` when you would rather follow the series forward.

## When to use

### Best For

- **Reproducible agent behaviour** where a dated checkpoint prevents a series update from shifting output
- **Agentic production work** that improved most between the preview and this release
- **Million-token context** made practical by compressed attention rather than only advertised
- **Terminal and tool workloads** where its vendor-reported Terminal Bench results are strongest
- **Shiftable batch traffic** that can run off-peak against the tiered rate

### Consider Alternatives When

- **General capability** is the priority, where its Artificial Analysis score sits mid-pack
- **Following the series** matters more than pinning, where undated `deepseek-v4-pro` tracks forward
- **Cheaper throughput** carries the workload, where `deepseek-v4-flash-0731` is the efficiency tier
- **Independently verified benchmarks** are required, since these results are vendor-reported

## Best for

- **Reproducible Agent Behaviour:** A dated checkpoint that a series update cannot shift
- **Agentic Production Work:** The area that improved most over the preview build
- **Million-Token Context:** Compressed attention making a 1.0M tokens window practical
- **Terminal And Tool Workloads:** Its strongest vendor-reported benchmark territory
- **Shiftable Batch Traffic:** Off-peak billing at half the peak rate

## Consider alternatives

- **General Capability:** Its Artificial Analysis score sits mid-pack against frontier models
- **Following The Series:** Undated `deepseek-v4-pro` tracks forward instead of pinning
- **Cheaper Throughput:** `deepseek-v4-flash-0731` is the efficiency tier of the same series
- **Verified Benchmarks:** Headline results are vendor-reported and not independently replicated

## Frequently asked questions

### What is DeepSeek V4 Pro 0813?

The general-availability checkpoint of DeepSeek's V4 Pro, released August 12, 2026 after a preview period of nearly four months. DeepSeek describes it as superseding the preview build with greatly enhanced agentic capabilities.

### Should I use DeepSeek V4 Pro 0813 or the undated deepseek-v4-pro?

Use DeepSeek V4 Pro 0813 when you need reproducible behaviour, since a dated slug pins one checkpoint. Use the undated slug when you would rather follow the series forward automatically.

### How is DeepSeek V4 Pro 0813 priced?

On a schedule with peak and off-peak rates, where off-peak bills at half the peak rate. Cache-hit input rose the most when the schedule changed. See the pricing panel on this page for current rates.

### What is the context window for DeepSeek V4 Pro 0813?

The context window is 1.0M tokens, with up to 1.0M tokens per response. Compressed attention cuts per-token compute and KV cache at long context, which is what makes the full window practical.

### How does DeepSeek V4 Pro 0813 perform on benchmarks?

Vendor-reported results are 87.9 on Terminal Bench 2.1, 74.1 on Toolathlon-Verified, and 80.6 on SWE-bench Verified. On the independent Artificial Analysis Intelligence Index it scores 53. Benchmark it on your own tasks before switching.

### Does DeepSeek V4 Pro 0813 support thinking mode?

Yes. Both thinking and non-thinking modes are available, alongside tool calls.

### Does DeepSeek V4 Pro 0813 support Zero Data Retention?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/deepseek-v4-pro-0813)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://api-docs.deepseek.com/)