The concept of Mixture-of-Experts predates deep learning, originating in a 1991 paper by Jacobs, Jordan, Nowlan, and Hinton. That original formulation used a dense-MoE structure where a gating network produced a weight for every expert. The 1991 paper demonstrated this through a vowel discrimination task. In that study, a system of competing experts plus a gating network outperformed a single monolithic network. The 2017 paper by Shazeer et al. bridged this to Transformers using top-k gating to allow only the most relevant experts to run. DeepSeek-V4-Pro utilizes this by having 256 routed experts and 1 shared expert per layer. The model selects the top 8 routed experts per token. This design allows the shared expert to absorb general knowledge while the routed experts handle narrower patterns. DeepSeek implemented a Hybrid Attention Architecture to combine Compressed Sparse Attention and Heavily Compressed Attention. This method reduces the KV cache footprint to 10 percent of what previous generations required. DeepSeek-V4-Pro activates 49 billion parameters during any single token generation despite its 1.6 trillion total parameter count. The model supports a 1-million token context window. Developers use the model for automated root cause analysis and multi-file code refactoring. DeepSeek-V4-Pro uses wide expert parallelism during training to spread the 256 routed experts across many GPUs. This prevents the model from becoming compute-bound.
Coding benchmarks favor OpenAI
DeepSeek-V4-Pro achieves a 55.4 percent score on the SWE-bench Pro evaluation, though it trails OpenAI’s GPT-5.5 which scores 58.6 percent and Anthropic’s Claude Opus 4.7 which reaches 64.3 percent in complex tasks. The model’s 1-million token context window provides a massive capacity to process entire code repositories. Full context length usage introduces latency. DeepSeek-V4-Pro scores 67.9 percent on Terminal-Bench 2.0, while GPT-5.5 reaches 82.7 percent on that same test. This performance gap shows the difficulty of matching closed-source leaders in complex command-line workflows. DeepSeek-V4-Pro excels at advanced reasoning, mathematics, and software engineering tasks through a two-stage post-training pipeline. This process involves Group Relative Policy Optimization and on-policy distillation to follow complex instructions. DeepSeek-R1 also shows competitive reasoning, scoring 97.3 percent on MATH-500 and 79.8 percent on AIME 2024. These scores exceed the 96.4 percent and 79.2 percent marks set by OpenAI o1-1217. DeepSeek-R1 scored 49.2 percent on SWE-bench Verified, but the May 28, 2025 update raised this score to 57.6 percent. OpenAI released GPT-5.5 on April 24, 2026, to improve agentic coding and computer use. GPT-5.5 claims to be a faster, sharper thinker that requires less human guidance. DeepSeek-V4 released in April 2026 and September 2026 focused on reasoning and long-context agentic coding.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| DeepSeek-V4-Pro | $2.00 | $4.00 |
| OpenAI o1 | $15.00 | $60.00 |
| DeepSeek-R1 | $0.55 | $2.19 |
The economic math of intelligence
DeepSeek-R1 output tokens cost $2.19 per million, which provides a 96 percent discount compared to OpenAI o1’s $60 per million. DeepSeek-V3 costs only $5.576 million to train using 2.788 million H800 GPU hours on 14.8 trillion tokens. DeepSeek’s researcher pool grew to 356 people across seven foundational papers. Fifty-three point five percent of researchers with known affiliations spent their entire careers at Chinese institutions. Most researchers with U.S. affiliations now work in China. DeepSeek’s pricing model destroys the premium margins of American labs. DeepSeek hit 96.88 million monthly active users by April 2025. This growth followed 173 million cumulative downloads since its launch. DeepSeek’s market share reached 56 percent in Belarus and 43 percent in Russia in January 2026. DeepSeek’s primary monetization channel is the API rather than a consumer subscription. OpenAI announced 800 million weekly active users for ChatGPT on October 6, 2025. OpenAI reached $12 billion in annualized revenue by mid-2025. You already know the basics of token costs. How long can the cost advantage hold?
