DeepSeek Ends V4 Pro's Preview Phase

DeepSeek launched the general availability version of V4 Pro on August 12, 2026, identified by the build code 0813, ending a preview period that lasted nearly four months. According to Unite.AI and Tech Times, the model now becomes the Chinese company's production flagship, no longer just a developer preview.

Technical Specs: a Large but Efficient Model

V4 Pro 0813 is a mixture-of-experts model with 1.6 trillion total parameters and roughly 49 billion active parameters per token, which keeps the computational cost of each response relatively low despite the model's overall size. It runs with a 1 million token context window and supports up to 384,000 output tokens, according to technical documentation released by the company.

Agent Focus: the Biggest Gains Came From Complex Tasks

The final version's most striking gains over the preview showed up precisely on agent benchmarks, tasks where the model uses tools, runs code and completes multi-step workflows without human intervention. The DeepSWE score jumped from 12.8 to 62.7 points, CyberGym went from 52.7 to 83.3, and Terminal Bench 2.1 climbed from 72.1 to 87.9, according to figures released by DeepSeek.

General Benchmark Performance Is Also Strong

Beyond the agent task gains, V4 Pro 0813 also posts competitive numbers on more traditional evaluations: 80.6% resolved on SWE-bench Verified, 90.1% pass rate on GPQA Diamond and 93.5% on LiveCodeBench, placing the model among the strongest publicly available options for coding and reasoning tasks.

Pricing Stays at Preview Levels for Now, but an Increase Is Coming

V4 Pro 0813's API keeps, for the moment, the same rates charged during preview: $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens, rates that remain accessible compared to competing frontier models. DeepSeek itself, however, warns on its pricing page that a significant increase is planned soon, with no date or magnitude disclosed.

Open Weights for the Final Version Have Not Been Released Yet

Despite DeepSeek's tradition of releasing open weights for its models, the 0813 version has not had its weights published yet: only the April preview version remains available for anyone who wants to run the model locally. There is no confirmed timeline for releasing the production version's weights.

Why It Matters for Brazilian Agencies and SMBs

For Brazilian marketing and customer service automation agencies, V4 Pro 0813 is another competitively priced frontier model option with a stated focus on agent tasks, the kind of capability used in chatbots that need to take action, not just answer questions. Since the promotional price could rise soon, it is worth testing the model now, while the low cost window lasts, but avoid locking clients into long term contracts based on a price DeepSeek itself has already warned will change.