OpenAI o-series
OpenAI pioneered the commercial reasoning-model category. o1 (2024) introduced hidden chain-of-thought with a "thinking" indicator. o3 (early 2025) showed dramatic gains on Frontier Math and ARC-AGI by scaling inference compute. o4-mini (April 2025) added native tool use during reasoning, allowing the model to call tools mid-thought. The exact backbone (dense vs MoE) is not public, but the reasoning behavior is purely a training + inference story.
Anthropic Claude with extended thinking
Claude 3.7 Sonnet (February 2025) introduced extended thinking with visible (but partially summarized) thinking blocks. Claude 4 (mid-2025) refined the budget controls. The same checkpoint serves both standard and extended-thinking modes via a configuration flag. Visible thinking quality has been a particular focus, with research on faithfulness published alongside the feature.
Google Gemini thinking modes
Gemini 2.0 Flash Thinking (December 2024) exposed visible reasoning. Gemini 2.5 Pro/Flash (2025) added explicit thinking-budget controls. Gemini 3.0 and 3.1 continue this with "Deep Think" modes that allocate larger budgets for harder tasks. Pricing reflects thinking-token cost; budgets are part of the API surface.
DeepSeek-R1 — the open-weight reasoning frontier
DeepSeek-R1 (January 2025) made reasoning open-weight. Built on the same 671B-A37B MoE backbone as V3, trained via SFT cold-start → GRPO RL → rejection sampling refinement. Visible <think> blocks. R1-Zero (the no-SFT variant) demonstrated emergent reasoning from pure RL. Together, R1 and R1-Zero are arguably the most important open-weight reasoning release of the era.
Qwen3 with dual-mode operation
Qwen3 series (2025) supports thinking and non-thinking modes in a single checkpoint. Same weights, two behaviors. Trained with GSPO (Generalized Simple Policy Optimization). The same model can serve both quick conversational queries and complex problems by toggling a flag, eliminating the need for separate fast and reasoning model deployments.
The cross-family pattern
Everyone reaches similar capability through similar training pipelines (SFT + RL with verifiable rewards) on similar backbones (decoder-only Transformers, dense or MoE). The differences live in the visibility choice, the pricing model, and the post-training data quality. The architecture, again, is not the differentiator.