What's new
now data Recomputed internal benchmark aggregates now data Attempted Chatbot Arena collection now data Updated speed measurements now data Pulled latest OpenRouter price index now data Validated official pricing snapshots 1h digest Apple launches HomeView, its first smart display for the home 1h jobs Jobs market snapshot refreshed 1h digest Published the 2026-10-13 daily digest 1h digest SpaceXSI bundles X Premium, Grok and Cursor into one subscription 1h digest The final Disrupt Stage lineup: Three days of conversations you won’t hear anywhere outside of TechCrunch Disrupt 2026 now data Recomputed internal benchmark aggregates now data Attempted Chatbot Arena collection now data Updated speed measurements now data Pulled latest OpenRouter price index now data Validated official pricing snapshots 1h digest Apple launches HomeView, its first smart display for the home 1h jobs Jobs market snapshot refreshed 1h digest Published the 2026-10-13 daily digest 1h digest SpaceXSI bundles X Premium, Grok and Cursor into one subscription 1h digest The final Disrupt Stage lineup: Three days of conversations you won’t hear anywhere outside of TechCrunch Disrupt 2026

AI Reasoning Models

Models ranked by reasoning, math, and logic benchmark performance. Reasoning models use extended "thinking" to solve complex multi-step problems.

Dedicated Reasoning Models

Models specifically designed for extended reasoning. This watchlist is ordered by release date, not by the suppressed legacy quality score.

Claude Sonnet 4.6 (batch)

Anthropic

Awaiting evidence $3.00

Default model; extended thinking Anthropic balanced frontier model.

Claude Opus 4.6 (batch)

Anthropic

Awaiting evidence $5.00

Most capable; 1M context beta; adaptive thinking Official Opus release retained for historical comparisons.

GPT-5.2 Pro (batch)

OpenAI

93.2 benchmark avg $10.50
Reasoning avg: 93.2

Premium reasoning tier Speed data hidden until it is refreshed from a current live measurement source.

DeepSeek V3.2

DeepSeek

Awaiting evidence $0.29

685B params (37B active) MoE; 90% off cache hits

O3 Pro

OpenAI

Awaiting evidence $20.00

Highest reasoning quality Speed data hidden until it is refreshed from a current live measurement source.

R1 0528

DeepSeek

Awaiting evidence $0.50

Updated R1 with improved reasoning Speed data hidden until it is refreshed from a current live measurement source.

Gemini 2.5 Flash (batch)

Google

Awaiting evidence $0.15

With thinking: $3.50 output

O3

OpenAI

Awaiting evidence $1.00

Reasoning model; 80% price cut from launch

O4 Mini

OpenAI

Awaiting evidence $0.55

Fast reasoning; excels at math/code

Gemini 2.5 Pro

Google

Awaiting evidence $1.25

Thinking model with 1M context

o3 Mini (batch)

OpenAI

Awaiting evidence $0.55

Cost-efficient reasoning

DeepSeek R1

DeepSeek

Awaiting evidence $0.70

Open-weight reasoning; CoT tokens billed as output

DeepSeek V3

DeepSeek

Awaiting evidence $0.32

Original V3; 671B params (37B active) Speed data hidden until it is refreshed from a current live measurement source.

o1

OpenAI

Awaiting evidence $15.00

First-gen reasoning model; chain-of-thought Speed data hidden until it is refreshed from a current live measurement source.

Phi 4

Microsoft

Awaiting evidence $0.07

14B params; excels at math reasoning; MIT license Speed data hidden until it is refreshed from a current live measurement source.

Reasoning Benchmark Rankings

All models ranked by average score across reasoning/math benchmarks (AIME 2025, ARC Challenge, GPQA Diamond, Humanity's Last Exam, LiveBench, MATH-500).

# Model Reasoning Avg
1 GPT-5.2 Pro (batch) Reasoning OpenAI 93.2
2 Claude Opus 4.5 (batch) Anthropic 87.0
3 Claude Sonnet 4.5 (batch) Anthropic 83.4
4 GPT-5.2 OpenAI 63.5

What are reasoning models?

Reasoning models (like OpenAI's o-series and DeepSeek R1) use extended "chain-of-thought" processing to work through complex problems step by step. They're particularly strong at:

  • Mathematics: Competition-level math problems (AIME, MATH-500)
  • Science: Graduate-level science questions (GPQA Diamond)
  • Coding: Complex software engineering tasks (SWE-Bench)
  • Logic: Multi-step logical deduction and constraint satisfaction

The trade-off is higher latency and cost — reasoning models "think" before responding, which takes longer but produces more accurate answers for hard problems.

Coding models → Full leaderboard → Compare head-to-head →