DeepSeek built the V3 base model for $5 million and trained the R1 reasoning model for $5.6 million. This low expenditure challenges the requirement for massive capital clusters to achieve frontier performance. DeepSeek used 2,000 Nvidia H800 GPUs to train its 671B-parameter model for $294,000. The training pipeline for R1 includes a cold-start supervised fine-tuning phase with thousands of structured chain-of-thought examples and a second reinforcement learning pass using the Generalized Policy Optimization algorithm. DeepSeek’s training involves pretraining on the DeepSeek-V3 base model, followed by reasoning-oriented reinforcement learning to optimize rewards for accuracy and language consistency. DeepSeek-R1-Zero also demonstrated that pure reinforcement learning can drive autonomous reasoning improvement, although it suffered from flaws in text generation and language mixing. DeepSeek’s Mixture-of-Experts architecture activates only 37 billion parameters per request, which reduces compute needs and allows R1 to match OpenAI o1 performance on AIME 2024 with a 79.8% score. DeepSeek R1 achieves 71.5 on GPQA diamond and 97.3% on MATH-500. DeepSeek V4 also provides a 1 million token context window. DeepSeek V4-Pro achieves 80.6% on SWE-bench.
Open weights challenge proprietary dominance
DeepSeek V4 Flash output tokens cost $0.28 per million, which is 107x cheaper than the $30.00 output cost for GPT-5.5. This price gap fuels the migration of enterprises toward alternatives after Anthropic overtook OpenAI in paid business adoption in April 2026. DeepSeek provides an MIT license that permits free distillation and commercial use. While OpenAI holds consumer mindshare, the high cost of proprietary reasoning models pushes developers toward open-weight alternatives. OpenAI shut down o1-pro on October 23, 2026. DeepSeek R1 provides the most significant disruption to the reasoning market in 2026. DeepSeek-R1 provides 96% cost savings compared to o1 pricing, which allows engineering teams to reduce operational expenditure by more than 20x when switching from proprietary o1 endpoints to DeepSeek-R1 or self-hosted distilled models. DeepSeek V4 Pro matches GPT-5.5 on isolated math and coding tasks at 18x to 107x lower cost. The intelligence required for GPT-4-level tasks now costs $0.055 via Qwen3.7 Flash, representing a 682x drop from March 2023. DeepSeek V4-Pro reaches 80.6% on SWE-bench.
| Model | Access Type | Input ($/1M) | Output ($/1M) |
|---|---|---|---|
| DeepSeek-R1 | Open-Weights | $0.55 | $2.19 |
| DeepSeek V4 Flash | Cloud API | $0.14 | $0.28 |
| GPT-5.5 | Proprietary | $5.00 | $30.00 |
Technical trade-offs and local viability
The MIT license allows users to run distilled models on local hardware. The DeepSeek-R1-Distill-Qwen-32B model outperforms OpenAI o1-mini and runs on local PCs via Ollama. DeepSeek offers distilled models of various sizes from 1.5B to 70B parameters based on Llama3 and Qwen2.5. For larger models, users may require hardware like the DGX Spark or Nvidia Neotron 3 Nano Omni to handle the necessary memory capacity. Quantization to 4-bit or 8-bit precision provides 4x inference speedups with negligible accuracy loss. DeepSeek provides transparency by streaming raw tokens inside a "thought" block for users to observe the cognitive path. This open approach contrasts with OpenAI, which hides raw reasoning chains behind safety filters. DeepSeek also uses context caching to reduce the cost of repetitive prefix tokens by up to 90%. You should evaluate these variables before deploying models in a production environment. DeepSeek-R1 lacks function calling, multi-turn capabilities, complex role-playing, and JSON output. The model also follows Chinese censorship regarding Tiananmen Square and Taiwan independence. Will the lack of agentic features prevent DeepSeek from dominating the enterprise market?
