DeepSeek shipped two pieces of news in the same 24-hour window today, and they pull in opposite directions.

The first: DeepSeek-V4-Pro-0813 is now generally available, out of the preview it entered in April. The official GA release comes with substantially improved agentic capabilities and MIT-licensed weights, meaning businesses can run it without ever touching DeepSeek’s API.

The second: starting August 16 at 16:00 UTC, DeepSeek is raising API prices for V4 Flash and V4 Pro by between 50% and 1,100%, depending on the model, token type, and time of day. A new peak and off-peak billing structure means costs vary by when your workloads run.

Both pieces of news matter for businesses building on AI. The model keeps getting better. The API is getting a lot more expensive.

What Launched Today

DeepSeek-V4-Pro-0813 is the production-ready version of the company’s flagship model, replacing the preview designation. The model packs 1.57 trillion total parameters and activates 49 billion per token via a Mixture-of-Experts architecture, positioning it directly against proprietary heavyweights like GPT-5.5, Gemini 3.1 Ultra, and Claude Opus.

Key specs in the GA release:

• 1 million token context window
• 384,000 token maximum output
• Thinking and non-thinking modes switchable via a reasoning_effort parameter (low, high, max)
• Responses API support and Codex integration added — both built explicitly for agentic workflows
• Available across web platform, mobile app, and API immediately

The agentic benchmark improvements are the headline. DeepSeek’s own published numbers show DeepSWE moving from 12.8 on the preview to 62.7 on the GA release — roughly a fivefold improvement on the primary benchmark for autonomous software engineering. Terminal Bench 2.1 went from 72.1 to 87.9. CyberGym from 52.7 to 83.3.

Independent verification is still underway as of writing, but the direction is consistent across the tests DeepSeek published.

DeepSeek Harness Also in Developer Preview

Alongside the model launch, DeepSeek also announced DeepSeek Harness, an open-source agent runtime framework now in developer preview under the MIT license.

The positioning is direct: an alternative to Claude Code and similar coding-agent environments. Harness is built around a modular architecture where every part of the agent runtime — tools, file handling, session management, long-running workflows — can be replaced with plugins. The company began forming the Harness team in May 2026 and moved to beta testing in August.

For developers, this matters because it means DeepSeek is now competing at the model layer AND the agent tooling layer. For businesses evaluating AI coding agents, there is now a third open-source option entering the space.

The Price Increase

This is where the news gets complicated for anyone currently building on DeepSeek’s API.

Effective August 16 at 16:00 UTC, DeepSeek is introducing peak and off-peak pricing for V4 Flash and V4 Pro. Peak hours are defined as 01:00–04:00 UTC and 06:00–10:00 UTC. Off-peak rates are half the peak rates.

The specific increases:

V4 Flash:

• Input tokens (cache miss): $0.14/M → $0.44/M at peak (214% increase)
• Output tokens: $0.28/M → $1.32/M at peak (471% increase)

V4 Pro:

• Input tokens (cache miss): $0.435/M → $1.32/M at peak (203% increase)

DeepSeek said the adjustment is intended to “allocate resources more reasonably” and shift developer workloads toward less congested periods. The off-peak discounts mean developers who can schedule batch jobs outside peak hours will see more modest actual increases.

For reference: developers who primarily use DeepSeek during off-peak hours and benefit from cached inputs will see increases well below the 1,100% headline figure. The worst case — peak hours, uncached inputs, output-heavy workloads — is the 1,100% scenario.

What This Means for Business

The “AI is getting cheap and will stay cheap” assumption is being tested. DeepSeek’s initial January 2026 launch disrupted the market by demonstrating that frontier-quality AI was achievable at a fraction of what OpenAI and Anthropic were charging. Competitors cut prices. The market adjusted to the idea that AI commodity economics applied.

Today’s announcement tells a different story. As DeepSeek’s models get more capable and demand increases, the company is raising prices toward a tiered structure similar to what the major cloud providers have used for years. The disruption may have been temporary pricing pressure rather than a permanent reset.

The open-weight advantage matters more now. DeepSeek-V4-Pro-0813 is MIT licensed. Any business with the infrastructure to run it can deploy it internally without paying DeepSeek’s API prices at all. The costs shift to compute, which you control, rather than API calls, which you don’t.

For enterprises already evaluating on-premise AI deployments, today’s model launch is a meaningful data point: this is now arguably the best open-weight model available, and running it yourself is a legitimate alternative to the API.

Agentic AI is advancing faster than the tooling for managing it. The fivefold improvement in DeepSWE scores between preview and GA, in roughly three months, suggests the capability frontier for AI agents is still moving quickly. Businesses that are not yet building with agentic AI are falling further behind — not because the technology is changing in ways that require constant rewiring, but because the useful benchmark for “what an agent can reliably do” keeps rising.

Budget for AI infrastructure volatility. Whether you are using DeepSeek, Anthropic, OpenAI, or Google, API pricing is not settled. Any business that has locked in an AI product or workflow built on a specific API’s cost structure should audit that dependency. The assumption that costs will remain flat — or continue to fall — is no longer safe.

If your team is still deciding whether to build AI into your operations, or if you are running AI workloads that have grown organically without formal cost tracking, now is the right time to do that analysis. The tools are better than ever. The pricing is more complex than it was six months ago.

---

Enterprise DNA helps business teams understand and act on AI developments like this one. If you want to assess how your current AI stack holds up against a changing model and pricing landscape, start with a discovery call.