DeepSeek drops V4-Pro with flexible reasoning and half-price off-peak APIs

If you're running automated AI agents in production, you know the drill: your API bills skyrocket, complex reasoning is painfully sluggish, and every edge-case failure burns actual cash.

DeepSeek just launched DeepSeek-V4-Pro into general availability, aiming straight at those production bottlenecks. It brings major upgrades for agentic workflows, flexible reasoning controls, and a pricing model that feels more like an electric utility.

A 1.7T beast with three throttle levels

Under the hood, DeepSeek-V4-Pro is a massive 1.7 trillion parameter Mixture-of-Experts (MoE) model. The raw unquantized model weighs in at 893 GB, and it introduces DeepSeek's new D-Spark speculative decoding to boost inference speeds.

Instead of leaving thought duration to chance, DeepSeek now gives you three distinct reasoning effort levels across both V4-Pro and V4-Flash: low for basic tasks, high for routine agent loops, and max for complex logic and coding.

Web and app users can test this right away via the "Expert Mode" toggle. For developers, model identifiers remain untouched in the API.

DeepSeek also added native support for OpenAI's Responses API with optimizations for Codex. If you've built your infrastructure around OpenAI endpoints, switching over is essentially a one-click affair.

Six cents a run and off-peak discounts

Across benchmarks like Humanity's Last Exam, Terminal Bench, and DeepSue, V4-Pro holds its own against open weights like GLM and Kimiko 3, alongside top-tier closed systems like Claude Opus.

Artificial Analysis's independent tracking places V4-Pro roughly tied with GLM 5.2 and slightly below Kimiko 3, lagging only a few points behind flagship closed models like GPT and Grok 4.6.

Where it completely breaks the curve is price. At an estimated 6 cents per task on Artificial Analysis, it drastically undercuts GPT-5.6, Kimiko 3, and Claude. On their intelligence-versus-cost scatter plot, DeepSeek sits virtually alone in the ideal top-left quadrant.

To twist the knife further, DeepSeek is introducing peak and off-peak API pricing starting August 16, 2026, at 16:00 UTC. Off-peak workloads get a full 50% discount, allowing scheduled batch jobs and agent runs to cost half as much.

Why it matters

We think this is one of the smartest product updates DeepSeek has pulled off. Instead of chasing marginal benchmark victories, they're aggressively optimizing for real-world production viability.

For developers, granular reasoning controls eliminate wasteful overthinking on trivial queries, while the off-peak rate makes heavy data pipelines and overnight agent tasks shockingly cheap. If you were holding off on deploying autonomous agents due to runaway API overhead, this changes the math.

The obvious catch is raw size. At 893 GB, self-hosting the full model requires a stack of serious enterprise accelerators. However, open-source maintainers like Unsloth are already working on compressed GGUF quantizations, so lightweight deployment won't stay out of reach for long.

For the wider market, this puts immense pressure on closed-source providers. When an open model matches elite agent benchmarks at six cents a task—and cuts that in half after hours—justifying triple-digit enterprise contracts becomes an increasingly tough sell.