DeepSeek-V4-Pro-0813

Frontier Scores, Open Weights, Quarter of the Price

DeepSeek-V4-Pro-0813 is the model that makes the rest of the leaderboard defend its pricing. The numbers first: 81.0 SWE-bench Verified, 80.3 Terminal-Bench, 176 tokens per second, a 1M token context window, and weights you can download. GPT-5.6 Sol leads it by 0.2 points of SWE-bench and charges 4.5 times as much per token. Whatever your position on open-source models, the score-per-dollar line on our leaderboard starts here.

What Open Weights Actually Buy You

The download is not just a license file - it changes your economics and your risk model:

• No per-token bill at scale. Self-hosted inference on your own GPUs has a fixed cost curve; past break-even volume, every token is free. If you serve more than a few hundred million tokens per month, run the math against the API prices in the table above.
• Data never leaves your infrastructure. For healthcare, legal, finance, and defense workloads, the ability to keep prompts and completions inside your own network is not a discount - it is a requirement no proprietary API can satisfy.
• Fine-tuning rights. You can specialize the model on your codebase or domain data. Every proprietary API is a black box you rent; this one is an asset you modify.

Benchmark Deep Dive, In Plain English

• SWE-bench 81.0 - Within one point of GPT-5.6 Sol (81.2) and ahead of every other open-weights model. For issue-resolution work, the capability gap between the free-weights option and the premium proprietary option has effectively closed.
• Terminal-Bench 80.3 - Live terminal work is where V4-Pro trails its closest rivals (Kimi K3 scores 85.7). If your agent lives in a shell all day, weigh that gap seriously.
• 176 tokens per second - The fastest model in our top tier. Speed compounds in agentic workloads: a 20-step agent loop finishes visibly faster, which means lower latency for users and lower compute cost per task.

Pricing Reality Check

The API pricing is $1.74 per million input and $3.48 per million output. The 10,000-request workload we use as a benchmark (12K in, 3K out each) costs about $313 - versus $1,170 on GPT-5.6 Sol and $2,170 on Claude Fable 5. And that is the expensive option, since the weights are available for self-hosting where marginal cost approaches zero.

Real Limitations

Math is the visible gap: MATH-500 at 52.8 trails Sol’s 64.5 and Kimi K3’s 59.8, which matters for quantitative pipelines. Web research (BrowseComp 41.8) is mid-pack. And self-hosting is a real operational commitment - you are trading API simplicity for GPU fleet management, quantization decisions, and version pinning. Teams without infrastructure staff should start with the API and graduate to hosted weights when volume justifies it. For a historical view of how the open-weights economics got here, see our Qwen 2.5 vs Llama 3.1 comparison.

Where to Go Next

• The value comparison: DeepSeek-V4-Pro vs GPT-5.6 Sol
• Open-weights landscape: Best LLM for Coding
• Where it ranks today: AI Model Leaderboard