DeepSeek-R1 drew global attention in January 2025 because DeepSeek released model weights and a technical report for a reasoning-oriented model that performed competitively on several self-reported evaluations. The event mattered, but many early comparisons mixed together four different questions: model quality, training compute, API price, and the capabilities of the ChatGPT product.

A rigorous comparison must name the exact systems. DeepSeek-R1 is a model family released in 2025. ChatGPT is a product whose available OpenAI models and tools change over time. Comparing “DeepSeek” with “ChatGPT” without model IDs, dates, settings, and tasks is therefore not reproducible.

What DeepSeek-R1 Actually Released

DeepSeek’s official R1 repository describes DeepSeek-R1 and R1-Zero as 671-billion-parameter mixture-of-experts models with 37 billion parameters activated per token and a 128K context length. The repository states that R1 is based on DeepSeek-V3-Base and reports evaluations against contemporaneous systems including OpenAI o1. It also includes smaller distilled models based on Qwen and Llama families.

The R1 code and model weights in the official repository are released under the MIT License, with inherited licenses applying to the distilled base models. “Open weights” is the precise claim. It does not mean that every training dataset, production service, moderation system, or development decision is fully disclosed.

DeepSeek’s benchmark results are useful evidence about the release, but they are vendor-reported measurements under specified prompting and sampling settings. They should not be converted into a universal statement that R1 is more accurate than an OpenAI product. Independent evaluation on representative prompts, with blind scoring and uncertainty estimates, is still required.

The $5.6 Million Claim: What It Did—and Did Not—Measure

The widely repeated cost figure came from the DeepSeek-V3 technical report, not from a full accounting of DeepSeek-R1. DeepSeek reported 2.664 million Nvidia H800 GPU-hours for V3 pretraining and approximately 2.788 million H800 GPU-hours for the complete reported V3 training run. Multiplying that compute by the report’s assumed rental rate of $2 per GPU-hour produces an estimate near $5.576 million.

That estimate does not include all research experiments, failed runs, data acquisition and preparation, salaries, infrastructure ownership, inference, or the separate post-training work that produced R1. It should be described as an estimated compute cost for the reported final V3 run—not the cost of the company, the model family, or the research program.

OpenAI’s GPT-4 technical report explicitly withheld architecture size, hardware, training compute, dataset construction, and similar implementation details. Claims that GPT-4o had 1.8 trillion parameters, was a fully dense model, used a particular number of GPUs, or cost a specific amount to train should not be presented as facts without a primary disclosure.

Price Is Endpoint-Specific and Time-Dependent

API prices are not a single “cost per million tokens.” Providers commonly charge different rates for uncached input, cached input, and output; some also vary prices by processing tier or time. The DeepSeek R1 launch page listed separate January 2025 rates for cache-hit input, cache-miss input, and output. Those historical launch prices should not be compared with a generic “OpenAI” number.

For a current evaluation, record the date and calculate:

  • the exact model and endpoint;
  • input, cached-input, reasoning, and output token volumes;
  • tool, search, storage, or image charges;
  • latency, retry rate, and concurrency limits;
  • human-review and failure-remediation cost; and
  • self-hosting hardware, operations, security, and utilization where applicable.

Provider pricing pages are live operational documents. Link to them rather than copying an undated table that will silently become obsolete.

Architecture Does Not Decide the Use Case

DeepSeek-V3 and the full R1 models use a mixture-of-experts architecture, activating a subset of parameters for each token. That can reduce computation relative to activating every parameter, but parameter activation alone does not determine end-to-end latency, energy use, factuality, safety, or total cost. Serving hardware, quantization, batching, context length, output length, software kernels, utilization, and workload shape also matter.

Likewise, architecture does not prove that a model is suitable for medical diagnosis, legal advice, financial decisions, cybersecurity operations, or regulatory compliance. High-stakes adoption requires domain-specific evaluation, access controls, privacy review, failure handling, human oversight, and evidence that matches the intended population and task.

Multimodality and Tools Must Be Compared at the Product Level

The January 2025 R1 release centered on text reasoning, while ChatGPT offered product features beyond a raw text-model endpoint. Those capabilities have evolved since publication. Compare the current products on the features actually required—such as image input, file analysis, browsing, tool calling, structured output, deployment controls, regional availability, and data-retention terms—rather than assigning one permanent label to either ecosystem.

What the Market Reaction Proved

DeepSeek-related news coincided with sharp market moves in several technology and energy-linked stocks in January 2025. A one-day market reaction records investor expectations at that moment; it does not establish how AI electricity demand, data-center construction, or semiconductor demand will evolve. Efficiency can lower the compute required for a fixed task while lower prices and new applications increase total usage. Long-term energy conclusions require measured deployment and demand data.

A Reproducible Evaluation Framework

  1. Name the candidates. Record exact model IDs, versions, endpoints, regions, and evaluation date.
  2. Freeze the task set. Use representative prompts and documents selected before viewing model outputs; separate development and final test sets.
  3. Define scoring. Measure task success, factuality, citation quality, safety failures, latency, and cost per successful task—not one synthetic “accuracy” score.
  4. Control the settings. Record system prompts, sampling parameters, tool access, context supplied, retry policy, and maximum output.
  5. Use blind review where possible. Randomize output order, use multiple qualified reviewers, and report disagreement and uncertainty.
  6. Test operations and governance. Evaluate privacy terms, data residency, access controls, audit logging, availability, rate limits, incident response, and model-change policy.
  7. Re-run periodically. Models, prices, and product features change; a 2025 result cannot be assumed to describe a 2026 service.

Bottom Line

DeepSeek-R1 was significant because it paired a capable reasoning model with released weights, a permissive license, and unusually detailed technical reporting. It did not prove that frontier AI could be built for $6 million in total, that OpenAI used a disclosed 1.8-trillion-parameter dense model, or that AI energy demand would fall. The defensible conclusion is narrower: DeepSeek demonstrated a competitive 2025 reasoning approach and changed the range of options available to evaluators. Choosing between current systems still requires a dated, task-specific, reproducible comparison.

Primary Sources