DeepSeek has released V4-Pro-0813 with major company-reported gains in coding and agent benchmarks, a one-million-token context window, MIT-licensed weights and native Responses API support. The upgrade also brings substantially higher API pricing.

DeepSeek just made V4-Pro official. On August 13, the Chinese AI lab took the model out of preview and rolled the full release, called DeepSeek-V4-Pro-0813, into its app, website, and API. It is not a quiet update. The company is reporting large gains on agent work-the kind of multi-step tasks where a model uses tools, runs code, works through repositories, and keeps going without constant hand-holding.

At the same time, DeepSeek is raising API prices and introducing peak and off-peak billing. The new rates take effect at 16:00 UTC on August 16, or 9:30 p.m. IST. That combination-a stronger agent model and a meaningful price increase-is what has people paying attention.

This is the same company that made waves earlier with capable models at unusually low prices. Now it is trying to move its flagship further up the performance ladder while retaining the open-weight approach that helped set it apart.

What Exactly Is V4-Pro?

V4-Pro is a mixture-of-experts model with 1.6 trillion total parameters, although only around 49 billion are activated for each token. That sparse design reduces the amount of computation required compared with a dense model of similar total size. It supports a one-million-token context window, while the official API allows a maximum output of 384,000 tokens.

The DeepSeek-V4-Pro-0813 weights are available on Hugging Face under an MIT licence. That gives organizations the option to inspect, adapt, and deploy the model on their own infrastructure. It does not, however, make self-hosting trivial. A 1.6-trillion-parameter model still demands serious accelerator capacity; DeepSeek’s own serving example uses a four-GPU GB300 node.

The official build retains the structure of the preview model and adds DeepSeek’s DSpark speculative-decoding module. On the API side, DeepSeek has introduced low, high, and max reasoning-effort settings, allowing developers to match the model’s thinking budget to the difficulty of the task. Simple work can stay on low effort, ordinary agent tasks can use high, and the most demanding workflows can use max.

DeepSeek’s API now also supports the OpenAI Responses API format natively and includes a one-click configuration route for Codex. For developers already working with Responses-style agent tooling, that should make testing or switching models less cumbersome.

The main story, though, is the agent performance. In DeepSeek’s published evaluations, V4-Pro-0813 improved from 72.1 to 87.9 on Terminal-Bench 2.1. Its DeepSWE score rose from 12.8 to 62.7, while CyberGym moved from 52.7 to 83.3 and Toolathlon-Verified increased from 55.9 to 74.1.

Those are substantial jumps, but they need the right framing. These are company-reported results. For the public code-agent benchmarks, DeepSeek used the minimal configuration of DeepSeek Harness with max reasoning effort, a temperature of 1.0, and top_p set to 0.95. Results obtained with another harness or configuration may differ. Because the model has only just been released, broader independent evidence from production deployments remains limited.

DeepSeek has also open-sourced DeepSeek Harness, the agent framework used in those code-agent evaluations. The project is available under an MIT licence, although it is still in developer preview and DeepSeek warns that compatibility-breaking changes should be expected.

What Makes It Stand Out

Plenty of models claim strong agent capabilities now. What makes V4-Pro interesting is the combination of capability, long-context efficiency, open weights, and comparatively low API pricing.

The long-context architecture is an important part of that. DeepSeek combines Compressed Sparse Attention and Heavily Compressed Attention in a hybrid design intended to reduce the cost of processing very long sequences. At a one-million-token context, the company says V4-Pro requires 27% of the single-token inference FLOPs and 10% of the KV cache used by DeepSeek-V3.2. That matters when an agent is working through a large repository, a long technical record, or a substantial collection of enterprise documents.

The MIT-licensed weights are another differentiator. Most frontier models from the largest US labs remain closed. Organizations with sufficient infrastructure can download V4-Pro, adapt it, and retain more control over deployment. For smaller teams, the official API will remain the more practical route.

Even after the price increase, V4-Pro remains considerably cheaper per text token than flagship models such as GPT-5.6 Sol and Claude Opus 5. That does not make the models directly equivalent, and token prices alone do not determine the cost of completing an agent task. Tool calls, retries, reasoning length, latency, and success rate all affect the final bill. Still, DeepSeek continues to occupy an attractive position for teams that want capable agent performance without paying the highest frontier-model rates.

The agent focus also feels practical rather than purely benchmark-driven. The reasoning-effort controls allow developers to spend more compute only when a task justifies it, while the Responses API and Codex integration reduce some of the friction involved in trying the model inside existing workflows.

None of this means V4-Pro leads every category. DeepSeek’s own results show a competitive model, not a universal winner. But for teams that care about coding agents, long-running tool loops, and deployment flexibility, V4-Pro looks like a serious option rather than another open-weight model merely trying to catch up.

The Pricing Shift Changes the Math

Until now, DeepSeek kept V4-Pro aggressively cheap. On the official API, one million cached input tokens cost $0.003625, uncached input cost $0.435, and output cost $0.87. That made high-volume agent runs unusually inexpensive, particularly when workloads achieved a good cache-hit rate.

The pricing changes at 16:00 UTC on August 16. DeepSeek is introducing two peak windows-01:00 to 04:00 UTC and 06:00 to 10:00 UTC-with all other hours treated as off-peak. Off-peak rates are half the corresponding peak rates.

For V4-Pro, one million cached input tokens will cost $0.022 off-peak and $0.044 at peak. Uncached input will cost $0.66 off-peak and $1.32 at peak, while output will cost $1.98 off-peak and $3.96 at peak.

That is a substantial increase from the old flat pricing. Cached input becomes roughly 6.1 times more expensive off-peak and 12.1 times more expensive at peak. Uncached input rises by around 1.5 times off-peak and three times at peak. Output pricing increases by approximately 2.3 times off-peak and 4.6 times at peak.

The comparison with V4-Flash is less dramatic. Under the new schedule, V4-Pro costs about three times as much as Flash for uncached input and output, with a little over a threefold difference for cached input. DeepSeek is therefore creating a clearer two-tier lineup: Flash for lighter or high-volume workloads and Pro for tasks where stronger performance may justify the additional cost. The exact rates and time windows are available on DeepSeek’s official pricing page.

The structure suggests a deliberate move away from competing almost entirely on price. DeepSeek appears to be positioning Pro as a higher-value product while retaining Flash as the lower-cost option. Whether Pro remains economical will depend heavily on the workload. Agent jobs run during the peak windows will become considerably more expensive, while scheduled or batch work can take advantage of the off-peak rates.

Compared with the flagship prices charged by OpenAI and Anthropic, V4-Pro still looks inexpensive on a per-token basis. But that is only one part of the calculation. A cheaper model that requires more retries may cost more per completed task, while a higher-priced model that finishes reliably in one run may prove more economical. Teams will need to compare completed-work cost, not merely the price of a million tokens.

How It Stacks Up Against the Competition

The field is crowded and moving quickly. On the closed side, Anthropic now has Claude Fable 5 and Claude Opus 5, while OpenAI’s GPT-5.6 Sol targets high-end coding and professional work. Google’s Gemini 3.7 Flash is pushing hard on agents, coding, speed, and multimodal workflows, and xAI’s Grok 4.6 is aimed at many of the same long-running engineering tasks.

Among open-weight and Chinese-origin competitors, Moonshot AI’s Kimi K3, Z.ai’s GLM-5.3, and Alibaba’s Qwen family are the other names likely to appear on enterprise evaluation lists.

DeepSeek’s launch table should not be read as a complete ranking of all those systems. It compares V4-Pro-0813 primarily with DeepSeek’s earlier models, Kimi K3, GLM-5.2, Claude Opus 4.8, and Claude Fable 5. Several newer models are not included, and benchmark results collected with different agent harnesses are not always directly comparable.

Within DeepSeek’s published table, the picture is mixed but credible. V4-Pro scores 87.9 on Terminal-Bench 2.1, narrowly behind Kimi K3 at 88.3 and Fable 5 at 88.0. On DeepSWE, its 62.7 trails Kimi K3 at 67.5 and Fable 5 at 70.0, but exceeds the reported scores for Opus 4.8 and GLM-5.2. V4-Pro leads that particular comparison on CyberGym with 83.3 and on the public portion of AutomationBench with 31.8.

That makes V4-Pro competitive across several agent-oriented tests without making it the outright leader everywhere. Its strongest argument is the package as a whole: strong reported agent performance, efficient million-token context, MIT-licensed weights, and an API price that remains below the most expensive closed models.

It is also worth noting that DeepSeek-V4-Flash-0731 outperformed the earlier V4-Pro preview on several agent benchmarks despite activating far fewer parameters. That shows how much of the recent improvement is coming from post-training, reasoning control, and the agent setup-not simply from increasing model size.

Real-world results will vary. Production agents depend on the tools, the harness, the prompts, the surrounding application logic, and the quality of the data they can access. A benchmark score is a useful signal, but it is not a substitute for testing the model on the exact workflow an organization intends to deploy.

Who Should Care and Why

Developers building coding agents or multi-step automation pipelines will want to test V4-Pro. Native Responses API support and the Codex configuration path reduce the effort required to put it inside an existing agent workflow. Organizations that already operate substantial AI infrastructure, or need the option to deploy models under their own control, also gain access to a capable MIT-licensed model.

For cost-sensitive projects, DeepSeek now offers a clearer internal choice. Flash can handle lighter or higher-volume work, while Pro can be reserved for tasks where additional reasoning and agent capability produce a measurable improvement. Flexible effort settings make that separation more useful because not every request has to run with the maximum thinking budget.

V4-Pro is less attractive for teams that need documented native image or video input through the public API. DeepSeek’s current API documentation presents V4-Pro as a text model and does not list multimodal input support. It may also be the wrong choice when the highest possible performance on a particular workflow matters more than deployment control or token price.

The timing is telling. The industry is deep into the agent phase of AI development. The models that matter increasingly are the ones that can remain coherent across many steps, use tools reliably, verify their own work, and recover when the first approach fails. DeepSeek’s decision to concentrate on those capabilities, keep the weights open, and then raise prices shows a company trying to mature its business without abandoning the qualities that first drew developers to it.

V4-Pro is not a complete rewrite of the AI landscape. It is a meaningful step forward for one of the industry’s most closely watched open-weight model developers. DeepSeek’s reported agent gains are substantial, the long-context efficiency work is technically important, and the MIT-licensed release remains a genuine differentiator.

The pricing change is equally important. It signals that the company wants to capture more value from its stronger model rather than remain known mainly as the inexpensive alternative. Even so, V4-Pro continues to undercut the flagship API rates charged by OpenAI and Anthropic, particularly outside DeepSeek’s peak windows.

Independent evaluations will matter over the coming weeks. So will reports from teams running the model against real repositories, enterprise documents, and operational workflows. For now, V4-Pro looks like a substantial and practical upgrade, but the strongest claims still come from DeepSeek’s own evaluation setup.

DeepSeek has already shown that it can move quickly and keep the economics interesting. With V4-Pro, it is testing whether it can also compete near the upper end of agent performance without giving up the openness and efficiency that made people pay attention in the first place. Real-world deployment will determine how well that bet pays off.