DeepSeek V4-Pro: Open Frontier Model at 1/10 the Cost

DeepSeek shipped V4-Pro-Max on Thursday under an MIT license. It scores 80.6% on SWE-bench Verified — the highest score any model has ever posted, open or closed. It also costs $1.10 per million output tokens. That is not a typo and it is not a promotional tier.

For context, Claude Opus 4.7 charges $15 per million output. Gemini 3.5 Pro charges $10. GPT-5.5 charges $12. DeepSeek is roughly 10x cheaper than the closed frontier, and it just outscored all three on the benchmark that actually matters for agentic coding work.

The numbers nobody wanted to see

V4-Pro-Max is a 1.6T-parameter mixture-of-experts model with 49B active parameters per token. Context window: one million. Training data cutoff: February 2026. Native tool-use, native multimodal, native 256-way parallel decoding.

On LiveCodeBench it scored 93.5 — beating GPT-5.5's 89.2 and Opus 4.7's 91.0. On AIME 2025 it solved 94 of 96. The model is not marginally competitive. It is at or above the frontier on every public benchmark that matters.

What this means if you run a real business

If you are an HVAC company paying $400 a month for Claude to draft service-ticket replies, you can do the same work for $40. If you are an agency running coding agents on Opus, your unit economics just improved by an order of magnitude — assuming you bother to switch.

Most won't switch. There is real friction. DeepSeek's API has latency variance the closed labs do not. Self-hosting V4-Pro-Max requires eight H200s minimum, which is not a Tuesday afternoon project. And enterprise procurement teams will spend six months arguing about whether a Chinese-origin model is allowed to touch customer data.

Where it shows up first

• Agentic workloads with high token consumption — the gap is biggest where you're burning tokens hardest
• Indie devs and bootstrapped SaaS — the people who actually compare prices instead of expensing them
• Open-source frameworks shipping with V4-Pro-Max as the default backend within 30 days
• Vertical AI startups quietly swapping out their Claude calls and pocketing the margin

The moat was never the model

Anthropic, OpenAI, and Google have been telling investors for two years that the moat is the model. The model is the product. The model is the differentiator. That story is now empirically false.

The real moat is distribution, integration, trust, and the boring middleware nobody wants to build. ChatGPT has 600M weekly actives. Claude is embedded in Cursor and every serious coding tool. Gemini ships inside Google Workspace. DeepSeek has a chat app and an API.

"Open source always wins eventually. The only question was when. The answer turned out to be this week."

— every developer on Hacker News, April 24

What to do about it

If you are building anything that calls an LLM API, stop signing 12-month contracts. The price floor is going to keep falling and the capability ceiling is going to keep rising and the lab that wins is not the one with the smartest researchers — it is the one that figures out how to make money when the model itself is free.

If you are an SMB owner using AI in your operations, this is good news. The thing you are paying for got 10x cheaper. Renegotiate. Switch providers. Or audit your current AI stack and accept that you are funding Anthropic's R&D out of inertia. Your call.