Drag each answer to line up with its matching prompt
OpenAI reasoning models (o-series, now GPT-5.x)
Google latency tier model with built-in thinking and opt-in thought summaries alongside fast multimodal chat
DeepSeek-R1
Proprietary slow thinking tier optimized for hard STEM with metered reasoning tokens
QwQ
Open RL first recipe with R1-Zero, cold start SFT, and public distillation checkpoints
Gemini Flash (thinking built in)
Anthropic product mode with bounded internal reasoning before the user visible reply
Claude extended thinking
Open weight reasoning model emphasizing math and code from Alibaba's Qwen line
Match vendors by openness and billing: OpenAI reasoning models metered closed STEM, DeepSeek open RL plus distill, QwQ Qwen math, Gemini fast summaries, Claude bounded thinking.
Every big AI lab now sells a model that thinks longer before answering, but they package it differently. OpenAI counts the hidden reasoning tokens on hard STEM and bills them like answer tokens. DeepSeek published the whole RL recipe and smaller copy models. Alibaba's QwQ focuses on open-weight math skill. Google blends quick chat with thinking summaries. Anthropic lets Claude think internally up to a budget before you see the reply.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Reasoning models stopped being a single OpenAI announcement and became a vendor map with different openness, billing, and trace policies.
The five offerings span closed metered frontier (OpenAI's reasoning models), open reproducible pipeline (DeepSeek-R1), open-weight specialist (QwQ), latency-tier multimodal (Gemini Flash), and bounded internal thinking (Claude).
OpenAI reasoning models: proprietary slow-thinking SKU
OpenAI's reasoning models (the o-series, now GPT-5.x) target hard STEM and multi-step logic with extra inference compute. OpenAI reports metered reasoning tokens separately in usage and bills them at the same output rate as visible completion tokens.
Training details are proprietary, but the product story is clear: pay for slow thinking when accuracy matters, not for every chat turn. Traces are generally not fully exposed to end users, which is relevant for safety monitoring discussions.
Procurement teams map vendors on four axes: peak benchmark, latency tier, openness/fine-tune rights, and thinking trace policy.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI's reasoning models report reasoning tokens separately in usage and bill them at the output-token rate
- DeepSeek-R1 and R1-Distill-Qwen-32B weights on Hugging Face enable self-hosted reasoning
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does trace visibility differ across these five offerings?
OpenAI/Anthropic often hide or bound thinking. Gemini summarizes. DeepSeek/QwQ may expose longer CoT in open weights. This affects safety monitoring.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Swapping open versus closed positioning: for example, calling DeepSeek the proprietary metered SKU or QwQ the Google product.
60 second bullets to scan on the way to the call.
Name OpenAI's reasoning models as proprietary metered slow-thinking
Place DeepSeek-R1 as open RL-first with distillations
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.