DeepSeek-V4-Pro-0813 Goes Live: 1.6T MoE With Breakthrough Agent Capabilities

Chinese AI lab DeepSeek has officially released DeepSeek-V4-Pro-0813, the general-availability build of its flagship mixture-of-experts (MoE) model. Released on August 13, 2026, the model steps out of a four-month preview phase with substantially enhanced agentic capabilities, a massive 1-million-token context window, and benchmark scores that place it among the top coding and reasoning models in the world — all while being priced at a fraction of what Western competitors charge.

The release marks DeepSeek’s most aggressive push yet to establish itself as the dominant open-weights alternative to frontier models from OpenAI, Anthropic, and Google. And it arrives at a moment of intensifying competition in the AI industry, where the gap between open and closed models continues to narrow.

What’s Inside: 1.6 Trillion Parameters, 49 Billion Active

DeepSeek-V4-Pro-0813 is a mixture-of-experts architecture with 1.6 trillion total parameters, of which approximately 49 billion are active per token. This sparse activation pattern — the hallmark of MoE design — allows the model to achieve frontier-level performance at a fraction of the inference cost of a dense model of equivalent size. The architecture routes each token through a subset of expert subnetworks, so while the model’s knowledge capacity rivals the largest dense systems, the compute required for any single forward pass remains manageable.

Key specifications include:

• Context window: 1,000,000 tokens (1M)
• Maximum output: 384,000 tokens
• Reasoning modes: Configurable, including a “Max Effort” thinking mode for complex multi-step tasks
• Input modalities: Text (multimodal extensions available through partner APIs)
• Weights: Open — available on Hugging Face under the deepseek-ai organization

The 1M context window is particularly notable. It allows developers to process entire codebases, lengthy documents, or extended conversation histories in a single prompt without chunking — a capability that has become increasingly important as agentic workflows demand broader context.

Benchmark Results: Coding Champion

The most striking aspect of the 0813 release is its benchmark performance, particularly in coding and agentic tasks:

• SWE-bench Verified: 80.6% — the highest score ever achieved by an open-weights model, tied with Gemini 3.1 Pro. This benchmark evaluates an AI’s ability to autonomously resolve real GitHub issues from popular Python repositories, requiring multi-file reasoning, debugging, and code generation.
• Aider benchmark: 71.6% — measuring performance on complex code editing tasks, placing V4-Pro ahead of Claude Opus 4 on this particular evaluation.
• Overall coding: Multiple community benchmarks report V4-Pro outperforming Claude 4 Opus in coding tasks while being up to 68× cheaper for typical workloads.
• Aggregate quality: On the vals.ai leaderboard, DeepSeek V4 Pro 0813 ranks second overall across 82 evaluated models with a composite score of 96.40%, within 0.60 points of the closed-source leader.

The SWE-bench Verified result is especially significant. At 80.6%, DeepSeek-V4-Pro-0813 demonstrates that it can autonomously resolve roughly four out of five real-world software engineering issues — a capability that until recently was exclusive to the most expensive closed models from OpenAI and Anthropic. For the open-source community, this represents a landmark: the best open-weights coding model available anywhere.

Agentic Capabilities: The Headline Feature

According to Reuters, DeepSeek explicitly emphasized that the V4-Pro-0813 build “greatly enhances agent capabilities.” The model is available through three channels: API, the DeepSeek app, and the web interface. The company highlighted that the new build is designed for multi-step reasoning, tool use, and autonomous task completion — the core building blocks of AI agent frameworks.

This focus on agentic performance aligns with a broader industry trend. Throughout 2026, the frontier of AI development has shifted from raw benchmark chasing toward practical agent deployments. Models are increasingly evaluated not just on their ability to answer questions, but on their capacity to navigate complex workflows, call external tools, maintain state across long interactions, and complete tasks that require dozens of sequential steps. DeepSeek’s investment in this area — combined with the massive context window — positions V4-Pro as a serious contender for developers building agent-based applications.

The Hacker News discussion around the release reflected cautious optimism. Commenters noted that while benchmark numbers are impressive, real-world agentic performance often diverges from controlled test scores. One developer observed that giving the model “a messy repo, tools and a task that takes 30+ steps” is the true test of whether the agent capabilities live up to the hype. Early reports from the community suggest the model handles these scenarios well, though comprehensive independent evaluations are still ongoing.

Pricing: Cheap, But Getting More Expensive

DeepSeek-V4-Pro-0813 is priced at $0.435 per million input tokens and $0.87 per million output tokens at standard rates. Cache-hit pricing drops to $0.003625 per million input tokens, reflecting the efficiency gains from MoE sparse activation. For comparison, Claude Opus 4 charges roughly $15–$75 per million tokens depending on the tier, making DeepSeek up to 68× cheaper for equivalent workloads.

However, in a move that sent ripples through the developer community, DeepSeek simultaneously announced significant API price increases on the same day. The new rates — ranging from 50% to 1,100% above current prices depending on model, token type, and time of use — will take effect at 16:00 UTC on August 16, 2026. The pricing structure introduces peak and off-peak tiers, with V4-Pro output tokens rising to $2 per million during off-peak periods and $4 during peak.

This dual strategy — releasing a dramatically more capable model while simultaneously raising prices — signals that DeepSeek is transitioning from a land-grab phase focused on user acquisition to a sustainability-focused model. The company had previously offered a 75% discount on V4-Pro when it first launched in April 2026, aggressively undercutting competitors. Now, with proven demand and benchmark-leading performance, DeepSeek appears confident that developers will pay more.

Reddit users reacted with mixed feelings. Some expressed frustration at the price hikes, particularly for high-volume applications. Others noted that even at the new rates, DeepSeek remains dramatically cheaper than alternatives — a 4× increase on a model that started at 1/68th the cost of Claude still leaves it at roughly 1/17th the price.

Availability and Access

DeepSeek-V4-Pro-0813 is available through multiple channels:

• DeepSeek API: The deepseek-v4-pro endpoint now points to the 0813 build
• DeepSeek Chat: Web and mobile app interfaces updated
• OpenRouter: Available for third-party integration at standard rates
• Hugging Face: Open weights under deepseek-ai/DeepSeek-V4-Pro-0813
• Partner platforms: Including Cursor, Aider, and various agent frameworks

The open-weights availability is a critical differentiator. Unlike closed models from OpenAI and Anthropic, developers can download, fine-tune, and self-host DeepSeek-V4-Pro — provided they have the hardware to run a 1.6T parameter model. The 49B active parameter count means inference is feasible on high-end multi-GPU setups, though running the full model requires substantial VRAM.

Industry Context and Implications

The DeepSeek-V4-Pro-0813 release comes amid a flurry of AI activity in August 2026. Google’s Gemini app just crossed 1 billion monthly active users. Anthropic began embedding invisible watermarks in Claude outputs. OpenAI is preparing to retire the o3 model. The competitive landscape has never been more intense.

DeepSeek’s strategy is clear: compete on capability while maintaining a dramatic cost advantage. The 80.6% SWE-bench score proves the model can hang with frontier systems on the hardest coding tasks. The open-weights approach builds loyalty among developers who value transparency and control. And the pricing — even after increases — remains disruptive.

The price hike, however, raises an important question about the long-term economics of open AI models. DeepSeek has been subsidized by its parent company and by China’s broader push for AI leadership. As the company moves toward financial sustainability, the gap between open and closed model costs may narrow. Whether DeepSeek can maintain its cost advantage while continuing to push the frontier remains to be seen.

For now, DeepSeek-V4-Pro-0813 stands as the most capable open-weights model available — and a compelling choice for any developer building AI applications in August 2026.