{"id":3931,"date":"2026-01-05T17:25:44","date_gmt":"2026-01-05T17:25:44","guid":{"rendered":"https:\/\/mazurly.com\/?p=3931"},"modified":"2026-01-05T17:45:57","modified_gmt":"2026-01-05T17:45:57","slug":"claude-3-5","status":"publish","type":"post","link":"https:\/\/mazurly.com\/fr\/claude-3-5\/","title":{"rendered":"Claude 3.5 Sonnet: Benchmarks, Features, and How It Compares to GPT-4o"},"content":{"rendered":"
<\/div>\n

Claude 3.5 Sonnet is Anthropic’s flagship model released June 21, 2024, matching GPT-4o’s reasoning while outperforming it on visual tasks and code at half the cost. It’s available free on Claude.ai and via API ($3\/million input tokens, $15\/million output).<\/p>\n\n\n\n

What Is Claude 3.5 Sonnet?<\/h2>\n\n\n\n

Claude 3.5 Sonnet<\/a> represents Anthropic’s first model in the Claude 3.5 family, delivering frontier-level reasoning at mid-tier pricing. The model operates at 2x the speed of Claude 3 Opus<\/a> while maintaining 200K token context window, making it ideal for complex reasoning, coding tasks, and visual analysis.<\/p>\n\n\n\n

Built on advanced constitutional AI principles, Claude 3.5 Sonnet excels at nuance, humor, and natural writing. It’s the company’s strongest vision model yet, surpassing Claude 3 Opus on standard benchmarks and competing directly with GPT-4o and Gemini 1.5 Pro.<\/p>\n\n\n\n

Key Features at a Glance<\/h2>\n\n\n\n
Feature<\/th>Details<\/th><\/tr><\/thead>
Context Window<\/strong><\/td>200K tokens (handles ~150-page documents)<\/td><\/tr>
Processing Speed<\/strong><\/td>2x faster than Claude 3 Opus<\/td><\/tr>
Vision Capabilities<\/strong><\/td>Text transcription, chart interpretation, diagram reasoning<\/td><\/tr>
Coding Ability<\/strong><\/td>64% agentic coding benchmark (vs. 38% for Claude 3 Opus)<\/td><\/tr>
Cost<\/strong><\/td>$3\/M input, $15\/M output tokens<\/td><\/tr>
Availability<\/strong><\/td>Claude.ai (free + limited), Pro subscription, API, Amazon Bedrock, Google Vertex AI<\/td><\/tr>
New Feature<\/strong><\/td>Artifacts (interactive code\/document workspace)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Claude 3.5 Sonnet Benchmarks: How It Stacks Up<\/h2>\n\n\n\n

Claude 3.5 Sonnet leads or ties on most reasoning benchmarks against GPT-4o and Gemini 1.5 Pro. The most dramatic improvements appear in visual reasoning, coding proficiency, and complex reasoning tasks.<\/p>\n\n\n\n

Benchmark Performance Comparison<\/h3>\n\n\n\n

Visual Math Reasoning (MathVista)<\/strong> Claude 3.5 Sonnet achieves 67.7% on visual math problems, outpacing GPT-4o (63.8%) and Gemini 1.5 Pro (63.9%). This reflects superior ability to extract data from charts, graphs, and visual equations.<\/p>\n\n\n\n

Science Diagrams (AI2D)<\/strong> All three models cluster around 94%+, with Claude 3.5 Sonnet at 94.7%, demonstrating strong visual understanding of scientific illustrations.<\/p>\n\n\n\n

Document Visual Q&A (ANLS)<\/strong> Claude 3.5 Sonnet scores 95.2%, beating GPT-4o (92.8%) and Gemini 1.5 Pro (93.1%) at extracting information from document images, receipts, and scanned text.<\/p>\n\n\n\n

Graduate-Level Reasoning (GPQA)<\/strong> Claude 3.5 Sonnet dominates at 92% (0-shot), significantly outperforming Claude 3 Opus (87%) and competitive with GPT-4o. This signals strong performance on research-level questions.<\/p>\n\n\n\n

Coding Proficiency (HumanEval)<\/strong> In agentic coding tasks (write\/edit\/execute code with tools), Claude 3.5 Sonnet solved 64% of problems versus Claude 3 Opus at 38%. It handles code migrations, legacy updates, and bug fixes with sophisticated reasoning.<\/p>\n\n\n\n

Head-to-Head: Claude 3.5 Sonnet vs. GPT-4o<\/h3>\n\n\n\n

Claude 3.5 Sonnet and GPT-4o are nearly matched on reasoning benchmarks (both ~92% GPQA), but differ strategically. Claude 3.5 Sonnet leads on visual tasks and costs less ($3\/$15 vs. OpenAI’s $5\/$15 per million tokens). GPT-4o holds slight advantages on math (76.6% vs. 71.1% on MATH) and maintains broader integration ecosystem.<\/p>\n\n\n\n

Winner by use case:<\/strong><\/p>\n\n\n\n