Claude 3 Sonnet

Claude 3 Opus

Quick answer: Claude 3 Opus (released March 4, 2024, claude-3-opus-20240229) was Anthropic's most capable model at launch — the first Claude to post near-human performance on graduate-level reasoning (GPQA), surpass GPT-4 Turbo on most benchmarks, and achieve near-perfect needle-in-a-haystack recall at 200K tokens. At $15/$75 per million tokens, it remains the reference point for understanding how far the field has moved since early 2024.

At a Glance

Where Claude 3 Opus leads

• Historically important: the model that first demonstrated Claude could match or exceed GPT-4 Turbo on knowledge, reasoning, and coding
• Near-perfect long-context recall (>99%) up to 200K tokens
• Still useful as a stable, well-characterised baseline for regression testing against 2024-era capabilities

Where it lags

• Fully superseded by Claude 3.5 Sonnet (better performance, 3× cheaper), Claude Opus 4.6, and the full 4.x series
• At $15/$75 per million tokens it is the most expensive model in the Claude 3 family — a price now outcompeted by far more capable successors
• Knowledge cutoff August 2023; max output only 4,096 tokens

Best for: historical comparisons, regression baselines, and teams that need to understand what "state of the art in early 2024" looked like.

What Claude 3 Opus Is

Claude 3 Opus was the flagship of Anthropic's Claude 3 family, released March 4, 2024 alongside Claude 3 Sonnet and Claude 3 Haiku. At launch it topped most published benchmarks for commercial models: 86.8% MMLU, 50.4% GPQA Diamond, and 60.1% MATH — all state of the art at the time. Anthropic's internal agentic coding evaluation had it solving 38% of tasks, which was the highest reported number before Claude 3.5 Sonnet raised the bar to 64% just three months later.

Its defining quality at release was a twofold improvement in factual accuracy over Claude 2.1 on difficult open-ended questions, and near-perfect (>99%) accuracy on the needle-in-a-haystack recall test at 200K tokens. The model also demonstrated what Anthropic described as "metacognition": in some cases, Opus recognised when a needle sentence had been artificially inserted into context rather than treating it as genuine source material.

From a Benchgen perspective, Claude 3 Opus is a 2024 baseline. Teams running benchmarks in early 2024 generated a lot of trajectory data against this model. Comparing those results to current-generation models using Benchgen's evaluation dimensions provides a concrete, quantified view of how much agent reliability has improved.

Specifications

Pricing

Prompt caching reduces input costs by up to 90%. Source: Anthropic pricing page.

Context Window

Claude 3 Opus has a 200,000-token context window — roughly 300 pages of text — with a maximum output of 4,096 tokens. The 4,096-token output cap is significantly lower than later models (Claude 3.5 Sonnet outputs up to 8,192; Opus 4.6 up to 128K). This limits its usefulness for tasks requiring long generated responses.

Public Benchmark Scores

Scores are reported by Anthropic and shown for context. They are not Benchgen measurements.

Claude 3 Opus vs Alternatives

Claude 3.5 Sonnet outperforms Claude 3 Opus on every benchmark listed above while costing 5× less. There is no current production use case where Claude 3 Opus is the optimal choice over its successors. (Rival scores from respective provider announcements; not Benchgen measurements.)

How Claude 3 Opus Performs on Real Agent Tasks

Claude 3 Opus was the first Claude model Anthropic ran through internal agentic coding evaluations: 38% task completion, at the time a high-water mark. That number was rapidly eclipsed — Claude 3.5 Sonnet hit 49% SWE-bench Verified just months later — but it established the evaluation methodology that Anthropic and the broader industry have refined since.

For Benchgen users, Claude 3 Opus trajectory data collected in early 2024 is valuable historical material. Running the same evaluation harness against it today and comparing to current models is one of the cleanest ways to quantify capability progress. Expect Tool-Call Accuracy and Goal Completion scores noticeably below what Sonnet 4 or Opus 4 achieve on the same tasks.

Use Claude 3 Opus via API

from anthropic import Anthropic

client = Anthropic(api_key="YOUR_API_KEY")

message = client.messages.create(
    model="claude-3-opus-20240229",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Analyse the technical risks in this architecture doc..."}],
)
print(message.content[0].text)

Related Models & Benchmarks

• Compare: Claude 3.5 Sonnet · Claude Opus 4.6 · GPT-4o
• Benchmarks: SWE-bench Verified · MMLU-Pro

Frequently Asked Questions

Specs and scores sourced from Anthropic's Claude 3 family announcement. Third-party benchmark scores attributed inline. Last updated 2026-06-19.

Claude 3 Opus

Claude 3 Opus is a large language model developed by Anthropic.