Claude 3.5 Sonnet Tops the Charts — Anthropic's Surprise Benchmark Leader

·
5 min read
·Digital AI News Staff

Anthropic's Claude 3.5 Sonnet surpasses GPT-4o on key reasoning benchmarks while cutting latency in half — and it ships with a completely redesigned Artifacts feature.

When Anthropic released Claude 3 Opus in March 2024, the consensus was that Opus sat at the top of the capability hierarchy while Sonnet served as the practical middle tier. Claude 3.5 Sonnet inverted that assumption entirely. Released in June 2024, it bested both GPT-4o and Gemini 1.5 Pro on a range of reasoning, coding, and graduate-level problem-solving benchmarks — including a 64% score on HumanEval coding and 71.1% on MMLU — while operating at roughly twice the speed of Opus at considerably lower cost.

The release arrived alongside a new Claude.ai feature called Artifacts, which gives the model a dedicated canvas to render code, SVGs, and interactive React components in real time. Artifacts quickly became the most-discussed feature in AI developer communities, enabling a new class of prompt that ends with "now make it interactive" or "turn this into a working dashboard." The canvas approach — later mimicked by ChatGPT's Canvas and Google's Workspace integrations — repositions the chat interface as a lightweight IDE rather than a conversational text box. Anthropic positioned this as a step toward what it calls "agentic work," where the model scaffolds complex multi-step outputs rather than just answering questions.

Under the hood, Anthropic emphasized improvements to Claude 3.5 Sonnet's instruction-following fidelity and its performance on nuanced writing tasks — two areas where the previous generation drew consistent criticism. The company's Constitutional AI training approach appears to produce models with noticeably different failure modes than RLHF-heavy competitors: Claude 3.5 Sonnet is less prone to sycophancy and more willing to push back on flawed premises, behaviors that enterprise buyers increasingly treat as feature requirements rather than quirks. Anthropic has committed to a 3.5 Haiku and a full 3.5 Opus release, with the latter expected to reset the top of the leaderboard once more.


Claude 3.5 Sonnet is available at claude.ai and via the Anthropic API. Artifacts is available to all free and Pro users.