DeepSeek V4-Pro-0813 Ships as GA Build With Agent Upgrades

Overview

DeepSeek published a news update on its official API documentation site on August 13, 2026, confirming that build "0813" of DeepSeek-V4-Pro is now the model's official general-availability release. According to DeepSeek, the update brings "major Agent upgrades with strong production gains." The 0813 build follows DeepSeek's earlier move of V4-Pro to general availability on July 24, 2026; this release replaces that build as the current production version rather than introducing a new model line. Alongside the model update, DeepSeek introduced configurable reasoning-effort levels, native support for OpenAI's Responses API format, and announced a tiered pricing structure, with off-peak rates half the peak rates, scheduled to take effect three days later at 16:00 UTC on August 16, 2026. DeepSeek-V4-Pro remains the company's largest, most capable reasoning and agentic model, and build 0813 is now the version developers reach by default through both the API and DeepSeek's app.

Architecture and New Capabilities

According to the model card DeepSeek published on Hugging Face, DeepSeek-V4-Pro is a mixture-of-experts (MoE) model with 1.6 trillion total parameters, of which 49 billion are active per token. It uses a hybrid attention design combining Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA), paired with Manifold-Constrained Hyper-Connections (mHC) and a Muon optimizer used during training. Pretraining drew on more than 32 trillion tokens of diverse, high-quality data. The model supports a 1-million-token context window with a maximum output length of 384,000 tokens, and runs on mixed FP4/FP8 precision: MoE expert parameters use FP4, while most other parameters use FP8. Weights are released under the MIT license.

The 0813 release notes center on what DeepSeek calls "major Agent upgrades with strong production gains," though the company's public notes for this build do not break that claim down into specific published benchmark deltas. The more concrete addition is a new reasoning_effort parameter, now configurable for both DeepSeek-V4-Pro and DeepSeek-V4-Flash, with three selectable levels: low, high, and max. This lets developers tune how much internal reasoning the model performs per request, trading latency and cost against depth of deliberation, without switching model names.

DeepSeek also added native support for OpenAI's Responses API format, which the company describes as "optimized for Codex with one-click setup." That framing targets developers running Codex-style, terminal-based coding agents, letting them point existing Responses-API tooling at DeepSeek's endpoints with minimal reconfiguration.

Access and Efficiency

DeepSeek-V4-Pro-0813 is reachable through the DeepSeek app and web interface via a mode the company calls "Expert Mode," and through the API using the existing, unchanged model names. Per Unite.AI's reporting, build 0813 became available on DeepSeek's own API and on OpenRouter around August 12-13, 2026, the latest checkpoint in an iteration cycle that has run since DeepSeek-V4 first appeared as a preview in April 2026 and reached general availability on July 24. Concurrency limits on the hosted API stand at 500 simultaneous requests for V4-Pro, versus 2,500 for the lighter-weight V4-Flash, according to the same reporting, a gap that reflects Pro's substantially larger active-parameter count.

Current API pricing, in effect at launch and ahead of the August 16 change, is as follows.

For long-context workloads, DeepSeek's own comparison against its earlier DeepSeek-V3.2 model, measured at a 1-million-token context, reports that V4-Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache. For teams processing large documents or codebases, that efficiency gain should translate into meaningfully lower serving cost at full context length, though DeepSeek has not published equivalent comparisons against competing frontier models.

Strengths and Limitations

Pros

1. Native reasoning_effort controls (low, high, max) give developers direct latency-versus-accuracy tuning without switching models.
2. MIT-licensed, 1.6T/49B-parameter open weights remain downloadable, with roughly 1.4 million downloads in the most recent month, according to Unite.AI.
3. A 1-million-token context window paired with DeepSeek's reported 27%-FLOPs/10%-KV-cache efficiency lowers the cost of long-context serving.
4. Native Responses API support with one-click Codex setup reduces integration friction for terminal-based coding agents.

Cons

1. DeepSeek's "major Agent upgrades" claim is not accompanied by specific published benchmark deltas for this build.
2. The tiered pricing does not take effect until 16:00 UTC on August 16, 2026, three days after this release, so today's flat rates are a moving target rather than a stable basis for cost planning.
3. Pro's 500-request concurrency ceiling is a fifth of Flash's 2,500, a real constraint for high-throughput production deployments.
4. On Terminal-Bench 2.0, V4-Pro-Max's reported 67.9% trails GPT-5.4's reported 75.1%.

Competitive Position and Outlook

DeepSeek's own Hugging Face model card lists the following Pro-Max mode benchmark scores.

Unite.AI's reporting adds further Pro-Max figures: 80.6% on SWE-bench Verified, described as roughly level with Google's Gemini-3.1-Pro and 0.2 points behind Anthropic's Claude Opus 4.6; 67.9% on Terminal-Bench 2.0, behind OpenAI's GPT-5.4 at 75.1%; and 90.1% on GPQA Diamond. Those figures place V4-Pro-0813 close to, but not uniformly ahead of, several other frontier models on coding and reasoning tasks. The MIT license and continued Hugging Face availability, following roughly 1.4 million downloads in the most recent month reported by Unite.AI, mean the open-weights ecosystem retains access to a model competitive with proprietary frontier systems on several benchmarks, without licensing restrictions on commercial use.

Conclusion

DeepSeek-V4-Pro-0813 is an incremental but concrete production update to DeepSeek's flagship model: configurable reasoning effort, native Responses API support for coding agents, and benchmark scores that sit near several proprietary frontier models on coding and reasoning tasks. It is most relevant to developers already building agentic or coding-focused applications on DeepSeek's API, and to teams evaluating open-weight alternatives to closed frontier models. Buyers should model the August 16, 2026 pricing change into any cost estimate: from 16:00 UTC that day, DeepSeek switches to peak and off-peak billing, with off-peak rates half the peak rates, replacing today's flat pricing.

Editor's Verdict

DeepSeek V4-Pro-0813 Ships as GA Build With Agent Upgrades earns a solid recommendation within the Other LLM space.

The strongest case for paying attention: configurable reasoning_effort levels (low, high, max) give fine-grained control over cost and reasoning depth. That alone raises the bar for what readers should expect in this space. Reinforcing that, MIT-licensed open weights remain downloadable, with roughly 1.4 million downloads in the most recent month — practical value rather than just headline appeal. The broader signal worth registering is straightforward: DeepSeek-V4-Pro-0813, published August 13, 2026, is the current production build of DeepSeek's flagship reasoning and agentic model, replacing the version that had held general-availability status since July 24, 2026. On the other side of the ledger, one constraint is real rather than a marketing footnote: the "major Agent upgrades" claim lacks specific published benchmark deltas for this build. It should factor into any serious decision. Layered on top of that, the tiered pricing takes effect 16:00 UTC on August 16, 2026, so today's flat rates are a moving target for cost planning — which narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.