#Reasoning Models

DeepSeek-R1 vs OpenAI o3: Test-Time Compute & Reinforcement Learning (2026)

Chain-of-Thought vs Native Reasoning Tokens: What Actually Helps

Prompted chain-of-thought and trained-in reasoning tokens look similar from the outside — both produce visible 'thinking' text — but they behave differently, and conflating them leads to wasted tokens.

Benchmarking Reasoning Models for Code Generation Tasks

Public coding benchmarks are a poor proxy for how a model will behave in your codebase. Here's the harness I built to benchmark reasoning models against tasks that look like actual work.

When to Use a Reasoning Model vs a Fine-Tuned Small Model

Two completely different ways to get a hard task right: make the model think harder at inference time, or make a smaller model already know the answer shape. Neither is universally correct.

Reasoning Models vs Fast Models: Choosing the Right Latency/Accuracy Tradeoff

Reasoning models think longer and get more right. Fast models answer instantly and get most things right, most of the time. Here's how I decide which one belongs in a given code path.