DeepSeek R1 changed the cost question for reasoning software
DeepSeek R1 made reasoning models available through hosted services and downloadable weights. The commercial question was which model could perform a specific job reliably and at an acceptable cost.
DeepSeek's release of R1 on 20 January 2025 gave software builders a new set of choices: a hosted reasoning service, downloadable model weights and smaller models trained from the larger model's outputs. The announcement attracted attention for its reported performance against established reasoning systems. For a company building a product, the commercial question was which part of that capability it needed to own, rent or adapt.
The technical paper, posted to arXiv two days later, explained the training methods behind those choices. It described reinforcement learning, a staged training process and experiments with distillation into smaller models. The findings suggested several ways to deliver useful reasoning capability, each with different consequences for latency, operating expense and control. They did not establish a single cost that would apply across products.
Two models answered different research questions
R1-Zero applied reinforcement learning directly to a pretrained base model without an initial supervised fine-tuning stage. The researchers reported the emergence of longer reasoning, checking and correction behaviours. They also described poor readability and language mixing. The experiment demonstrated a training approach that still depended on pretraining data and other substantial inputs.
R1 added cold-start examples and a broader sequence of supervised fine-tuning and reinforcement learning. Supervised data therefore remained central to R1, even though the R1-Zero experiment omitted that initial stage. A product team could learn from the method without assuming that reproducing the full model was an economical use of its own capital.
Distillation offered a separate development route
The release included six smaller dense models based on Qwen and Llama, ranging from 1.5 billion to 70 billion parameters. Distillation used outputs from the stronger model to train a smaller one. In one experiment, the paper compared a Qwen 32B model trained through reinforcement learning with a 32B model distilled from R1. The distilled model performed better on the reported reasoning benchmarks, including AIME 2024.
| Period or category | % AIME 2024 pass@1 |
|---|---|
| R1-Zero-Qwen-32B: reinforcement learning | |
| R1-Distill-Qwen-32B: distillation |
DeepSeek R1 paper, Table 6. Reported AIME 2024 pass@1 results for two training approaches. The comparison does not establish equal training costs or performance on customer tasks.
Source2
The comparison tested two ways of developing a model of similar size. It did not measure equal training budgets, production latency or the cost of running a customer's workload. For a startup serving a defined task, it supported testing whether a smaller model could retain enough task performance to simplify deployment. A reliable specialist could be sufficient for the product.
Downloadable weights changed the deployment choice
The repository available at release documented local serving alongside the hosted chat service and API. Its licence section allowed commercial use of R1 and identified the underlying Qwen and Llama licences for distilled variants. That gave businesses more room to evaluate deployment options, while making the exact model and its provenance relevant to procurement. Buyers needed to check the licence attached to the specific release they planned to use.
Scroll to view all columns
| Comparison | R1-Distill-Qwen-32B | R1-Distill-Llama-70B |
|---|---|---|
| AIME 2024 · pass@1 | 72.6% | 70.0% |
| MATH-500 · pass@1 | 94.3% | 94.5% |
| GPQA Diamond · pass@1 | 62.1% | 65.2% |
| LiveCodeBench · pass@1 | 57.2% | 57.5% |
Results reported by DeepSeek at launch in its Distilled Model Evaluation table. The models differ in both size and base family, so the comparison does not isolate the effect of parameter count. These are the developer's results, not an independent replication. Rankings vary by task, supporting tests on the intended workload rather than assumptions about latency or cost based on model size alone.
Source3
Running a model within a chosen environment could help a supplier meet a customer's data handling or availability requirements. It also transferred work to the operator: capacity planning, model serving, upgrades, monitoring and incident response. A lightly used service might find a hosted API more economical than keeping dedicated hardware available. A predictable, heavily used workload could produce a different answer. Access to weights expanded the options, leaving the operator to compare their costs under its own workload.
Reasoning tokens were only one part of the bill
A model that takes longer to work through a problem can improve some answers while consuming more inference resources. The paper described response length increasing during R1-Zero's training. In a commercial product, the question becomes whether additional computation reduces costly mistakes or simply delays a response. An overnight analysis and a customer waiting at a checkout impose very different latency constraints.
An evaluation should therefore follow the complete task. If an answer needs a retry, a second model or a person's correction, those costs belong alongside the initial inference charge. A lower token price can coexist with a higher cost per accepted result. Conversely, a more expensive reasoning call can be worthwhile if it resolves an exception that previously required substantial manual work. These costs need to be measured on the customer's actual workload.
Product teams still had integration work to do
DeepSeek's paper acknowledged that R1 lagged its V3 model on tasks including function calling, interactions across several turns and JSON output. It also described prompt sensitivity and limited gains on software engineering benchmarks. Those limitations were particularly relevant to enterprise applications, where a correct calculation is only useful if it reaches the right record, respects permissions and produces a format the next system can use.
For investors, the release put greater weight on how a company applied a model to a working business process. A software company still needed distribution, evaluation on customer tasks, trustworthy integrations and evidence that customers used the output. Easier access to capable models could lower an input cost for many competitors at once. Companies that consistently turned that shared capability into completed work would have a stronger basis for competing.
Investment implications
- The released R1 model used supervised training stages as well as reinforcement learning; the R1-Zero experiment followed a different path.
- A smaller distilled model could be commercially useful without matching the largest model on every task.
- The relevant operating measure was cost per accepted result, including retries and review.
Sources
- DeepSeek-R1 Release (opens in a new tab)
DeepSeek
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, version 1 (opens in a new tab)
DeepSeek-AI / arXiv
- DeepSeek-R1 repository: launch-day README, including local deployment and licence terms (opens in a new tab)
DeepSeek-AI
Published
