2024Model

Claude 3 Series Released

Anthropic releases Claude 3 series, with Opus version surpassing GPT-4 on multiple tests.

#llm #claude #anthropic #multimodal #reasoning

Overview

In March 2024, Anthropic released the Claude 3 series, including Haiku, Sonnet, and Opus versions. Claude 3 Opus surpassed GPT-4 on multiple benchmarks, demonstrating strong reasoning, math, and coding capabilities. The Claude 3 series natively supports multimodal input, able to process images and documents. Opus particularly excels at complex tasks and long contexts (supporting 200K tokens), performing remarkably on graduate-level reasoning tests. This release marked a new stage in AI competition, with multiple companies fiercely competing in top-tier models.

Study Notes

Background, mechanics, impact, and key takeaways in one read

About 2 min read

Background & Context

On March 4, 2024, Anthropic released the Claude 3 family: Haiku (fast), Sonnet (balanced), and Opus (strongest). It was Claude's first truly multimodal generation, natively supporting image inputs such as charts, photos, handwriting, and technical diagrams, while raising context to 200K tokens across the family. On official benchmarks, Opus surpassed GPT-4 on several reasoning, math, and coding evaluations, entering the first tier on graduate reasoning (GPQA), undergraduate knowledge (MMLU), and code (HumanEval). Unlike OpenAI's release rhythm, Anthropic made the launch a paradigm of capability ladder plus safety reporting, attaching model cards and safety evaluations to each model. Claude 3 marked the frontier race entering a three-track phase of multimodality, long context, and safety governance.

Key Points

1Released March 4, 2024, with Haiku, Sonnet, and Opus forming a clear capability ladder.
2First native image inputs: charts, photos, handwriting, and technical diagrams.
3200K-token context across the family, handling hundreds of pages or whole codebases at once.
4Opus surpassed GPT-4 on several benchmarks, entering the first tier.
5Launches shipped with model cards and safety evaluations, making governance part of the product process.

How It Works

Claude 3's three engineering highlights. First, vision encoding: images are split into patches projected into visual tokens interleaved with text tokens sharing one Transformer context, letting the model cite charts and prose in a single inference. Second, long-context training: a 200K window required positional-encoding extrapolation (adjusted RoPE base frequencies) and curriculum training on long documents, and Anthropic disclosed needle-in-a-haystack results showing the model could find an implanted irrelevant sentence within 200K tokens, proving long context was not a paper parameter. Third, the capability-safety dual track: each tier corresponded to a Responsible Scaling Policy level, with red-teaming and third-party evaluations before deployment. This combination of tiers, evaluations, and model cards later became the de facto industry standard for launches.

Impact & Legacy

Claude 3 was the starting gun of the 2024 frontier reshuffle: it proved GPT-4's lead could be erased within a year, after which 2024's main theme became rapid handovers among GPT-4o, Claude 3.5, and Gemini 1.5. It also entrenched two industry standards: 200K context as flagship default, with Gemini following at 1M, and safety evaluations attached to launches. In the market, Claude 3 won Anthropic key enterprise and developer share, especially in coding, setting up Claude becoming many developers' first choice by late 2024. Criticism focused on the marketing nature of vendor-run benchmarks, prompt-format differences, and score-gaming risk, which raised the authority of independent evaluations like the LMSYS arena.

Key Takeaways

The right way to do multimodality is visual and text tokens sharing one Transformer context.
Long context must be proven with needle-in-a-haystack tests, not just window size.
Model cards, safety evaluations, and RSP tiers embed governance into the release process itself.
Frontier leads now have a half-life under a year, so release cadence is itself competitiveness.

Glossary

Needle-in-a-Haystack Test大海捞针测试

An evaluation that plants a specific fact deep in a long context and checks whether the model retrieves it.

Responsible Scaling Policy (RSP)负责任扩展政策

A governance framework, used by Anthropic and others, that ties safety measures to capability tiers, institutionalizing staged deployment.

Model Card模型卡

A document released with a model describing intended use, capability limits, evaluation results, and known risks.

Related Events

2023