On January 20th, DeepSeek released R1, an open-weight reasoning model that benchmarks roughly on par with OpenAI's o1 at about 30x lower cost. The weights are MIT-licensed, the technical report is published, and the model can be run on infrastructure that doesn't require US export-controlled GPUs. The engineering twitter has been thoroughly worked up about it.
| What is being compared | DeepSeek-R1 | OpenAI o1 |
|---|---|---|
| Cost | About 30x lower | Frontier pricing. The baseline the 30x is measured against |
| Reasoning quality | Benchmarks roughly on par | The model R1 is being compared against |
| Weights | Open-weight, MIT-licensed | Closed-source |
| Hardware | Runs on infrastructure that doesn’t require US export-controlled GPUs | Not stated |
| What the benchmark misses | Not a substitute where these matter | Tool-use polish, multimodal coverage, agentic harness maturity, alignment investments |
I want to write the boring version of this take: not "is the model good" (it's good), not "what does this mean for the AI race" (we'll find out), but what it changes for product managers shipping AI features in 2025.
The capability floor just moved up, quietly
For the past eighteen months the AI PM playbook has assumed something like a duopoly. You picked OpenAI or Anthropic, paid the closed-source rates, accepted the rate limits, and built features around the assumption that "good enough reasoning" was a frontier capability with frontier pricing.
R1 doesn't invalidate that picture, but it does add a floor. The minimum cost of a credible reasoning model is now an order of magnitude lower than it was two weeks ago. That doesn't mean every team will switch (latency, ops burden, support paths all matter). It means the BATNA changed. Frontier-model pricing is now negotiable in a way it wasn't.
I don't think this matters much for premium B2B AI products in the short term. It matters a lot for the cost-sensitive use cases I've been quietly bucketing as "wait until prices fall". Internal tools, customer-support copilots, classification pipelines. Those are suddenly economic.
What I'd change in my own roadmap this week
Three concrete moves:
Re-cost any AI feature still on the proposal stack. If a feature looked uneconomical at OpenAI rates a quarter ago, run the math again with open-weight reasoning costs in mind. Some of the features that died in cost reviews are now alive. Worth surfacing those before someone else does.
Re-evaluate features where you absorbed a "thinking cost" you'd rather pass on. If you're paying for o1-style reasoning to power something that ships free to your users, you have a margin problem masquerading as a model problem. The open-weight option changes the math without changing the user-facing capability.
Watch the eval landscape. Open weights mean third parties will be fine-tuning, distilling, and benchmarking R1 publicly. The lifecycle of "model released → public eval set forms → quality settles → adoption decision is rational" used to take quarters. With R1, it's been compressing into weeks.
What this isn't
This isn't the end of frontier-model premiums. The best closed-source models still have advantages: tool-use polish, multimodal coverage, agentic harness maturity, alignment investments that don't show up in a single-shot benchmark. If your product depends on those, R1 isn't a substitute.
It also isn't a green light to switch providers casually. Provider switching has a real cost (eval set portability, prompt tuning, output formatting drift, ops procedures). The right question isn't "should I move to R1?". It's "should I be running an eval against R1 as part of my next quarterly model review?". For most teams the answer to the second question is yes.
The PM-specific muscle this exercises
The capability commoditization story has been told a thousand times in software (databases, monitoring, deployment, etc.). The pattern is always the same. A category that was a paid frontier becomes a default commodity, and the value moves up the stack. The teams that ride the transition well are the ones that don't religiously stay on the incumbent.
For AI features in 2025, the value is going to move from "which model" to "which evals, which retrieval, which fallback, which UX". Get good at those. They're the part the model commoditization can't touch.
The boring version of the DeepSeek-R1 take: a commodity layer just got cheaper. Your roadmap should notice.
Further reading
- DeepSeek-R1 release announcement (Jan 20, 2025): DeepSeek's own launch post with pricing and benchmark numbers.
- DeepSeek-R1 on Hugging Face: MIT-licensed weights and model card.
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv 2501.12948): the technical report.
- DeepSeek-R1 GitHub: code and inference scripts.
- Learning to Reason with LLMs (OpenAI): the o1 announcement R1 is being compared against.
