A Chinese lab matched o1-preview on maths. Reasoning is not a monopoly
DeepSeek released R1-Lite-Preview on 20 November 2024, claiming o1-preview-level results on AIME and MATH. The lesson is about pace, not geopolitics.
On 20 November 2024 the Chinese lab DeepSeek released R1-Lite-Preview, a reasoning model it says reaches o1-preview-level performance on the AIME and MATH benchmarks, and which shows its chain of thought while working.
OpenAI announced o1 in September. Ten weeks later there is a public competitor with comparable claimed results on those tests. That interval is the story.
What is worth noting
The lead time on a new capability is now measured in weeks. Not years, not quarters. Whatever a frontier lab demonstrates, assume something comparable will be broadly available within a few months, often at a fraction of the price.
Benchmarks are not deployment. Matching a score on mathematics competition problems says little about drafting a Swiss employment contract, reading a handwritten delivery note, or handling German dialect in a voice memo. Be sceptical of the leap from leaderboard to workplace.
The visible reasoning is genuinely useful. A model that shows its working lets a human check the path, not just the answer. For anything an auditor might later question, that is more valuable than a marginal accuracy gain.
The part that needs care
Using a hosted service means your prompts go to that operator’s infrastructure under that operator’s jurisdiction and terms. For a Swiss company with client data, the location and legal framework are not a footnote; they are the decision.
This applies to every provider, not only this one. The right posture is consistent: know where the inference happens and under what agreement, for whichever model you use.
The takeaway for a planner
Stop building three-year plans around one lab’s roadmap. In a field where a capability gap closes in ten weeks, the durable asset is not your choice of model. It is your ability to change it: clean data, a documented process, and contracts short enough to move.
Companies that built for flexibility in 2024 will spend 2025 choosing. Companies that built for one provider will spend it migrating.