DeepSeek and open-source labs challenge tech giants with low-cost AI reasoning models

What happened with open-source reasoning models

AI research labs and open-source developers have released open-weight reasoning models that perform on par with proprietary systems on complex mathematics and coding benchmarks, at a fraction of the cost.

The shift centres on reinforcement learning techniques applied during post-training, combined with test-time compute—allowing models to generate internal reasoning steps before producing a final answer. Rather than relying solely on scaling up pre-training dataset sizes and cluster hardware, developers are using targeted training methods to teach models how to verify their own logic.

DeepSeek, an AI research company based in Hangzhou, published technical details showing that its reasoning architectures achieved high scores on competitive benchmarks such as AIME and MATH while spending significantly less on GPU compute than Silicon Valley competitors.

Why test-time compute changes model performance

Traditional large language models predict the next word in a sequence based on statistical patterns learned during training. Reasoning models, by contrast, allocate extra processing time during inference to explore multiple paths, check intermediate steps and correct errors.

This approach changes the fundamental economics of AI development. Increasing compute during the response phase allows smaller models to outperform much larger base models on tasks requiring multi-step logic. Research groups at Stanford University and independent labs have published findings confirming that test-time scaling can compensate for smaller pre-training parameters.

For developers, open-weight implementations mean complex reasoning capabilities can now be hosted locally or run on public cloud infrastructure without relying on restrictive proprietary APIs.

How proprietary labs are responding

Proprietary developers including OpenAI, Google and Anthropic are adjusting their research roadmaps to preserve their lead. OpenAI launched its own reasoning models, o1 and o3, which rely heavily on chain-of-thought processing during inference.

However, the rapid arrival of open-source equivalents has created price pressure across the software industry. API pricing for flagship reasoning endpoints has dropped sharply over recent months as providers compete for enterprise developers.

Industry analysts note that while proprietary labs retain an advantage in raw compute scale and multimodal capabilities, the performance gap for pure logic and code generation has narrowed faster than expected.

What comes next for AI development

The focus of AI research is moving toward autonomous agents capable of executing long-horizon tasks, such as multi-file software engineering, scientific data analysis and automated research pipelines.

Combining reasoning models with tool execution allows AI agents to plan actions, run tests in sandbox environments and iterate until a goal is met. Researchers at leading institutions indicate that the next bottleneck lies in reliable memory management and benchmark design, as current evaluation suites quickly become saturated.

Whether open-source models can maintain this pace remains unconfirmed, as future breakthroughs may depend on synthetic data generation techniques and custom hardware access that require substantial capital investment.