"Thinking model" became a mainstream concept in 2025. Nearly every leading AI company has launched its own reasoning model, all sharing the same core selling point: the model "thinks" internally before responding.
Approaches by Company
OpenAI o-series: The first to productize the "thinking" concept. Iterating from o1 to o3, it follows an "explicit reasoning chain + hidden thought process" approach. Users see only the final answer; the intermediate reasoning is not exposed.
DeepSeek R1: The representative open-source reasoning model. Driven by reinforcement learning, its thought process is transparent (the full Chain-of-Thought is visible). The 32B distilled version matches o1-mini on several benchmarks, with a massive cost advantage.
Gemini 2.5 Pro: Google's thinking model, which responds after internal reasoning. It excels in math and scientific reasoning (AIME 2024: 92%), with multimodal reasoning as a key differentiator.
Claude Opus 4.6: Adaptive reasoning — the model automatically determines the depth of thinking needed. Four adjustable levels prevent wasted compute.
Core Trade-off
All reasoning models face the same challenge: reasoning depth vs. response speed.
Deeper thinking yields more accurate answers, but increases wait time and cost. For simple daily conversations, enabling deep reasoning is simply a waste of money and time.
Each company's solution differs slightly:
- OpenAI: Offers different tiers of models (o1-mini to o3-pro)
- DeepSeek V3.1: Dual think/non-think modes within the same model
- Anthropic: Adaptive thinking — the model decides how deep to go
- Qwen3: Also a mixing mode, handling both thinking and non-thinking in one model
Open Source vs. Closed Source
Reasoning capability was once considered a moat for closed-source models. The emergence of R1 shattered that — open-source reasoning models have caught up with the flagship products of closed-source models. This means pure reasoning ability is increasingly difficult to sustain as a paid barrier.
The next battleground has shifted from "who can think better" to "who can think faster and cheaper."
Sources: CocoLoop, official model releases, The Decoder comparative analysis