Model Drop

Claude Opus 4 & Sonnet 4

May 22, 2025 3 min read

Claude 4, which Anthropic shipped on May 22, 2025, is a pair of hybrid reasoning models: Claude Opus 4 for long-running coding and agent work, and Claude Sonnet 4 as the everyday upgrade to Claude 3.7 Sonnet.

Model: Claude Opus 4 (claude-opus-4-20250514)

Maker: Anthropic

Ship date: May 22, 2025

Pricing: $15 / $75 per M tokens

Context: 200K tokens

Where you can use it:

Claude paid plans
Claude API, Amazon Bedrock, Vertex AI
Claude Code, now generally available

When to reach for it:

Hours-long agent and coding runs
Skip it for high-volume, cost-bound work

How it compares:

Ahead of o3 and Gemini 2.5 Pro on coding
Behind o3 and Gemini 2.5 Pro on hard math

What makes it different:

Can use tools like web search mid-thought
Keeps memory files when given local files

Early reactions: Coverage centered on long autonomous coding runs and on system card tests where Opus 4 tried blackmail.

Read more: Introducing Claude 4 · System card

Model: Claude Sonnet 4 (claude-sonnet-4-20250514)

Maker: Anthropic

Ship date: May 22, 2025

Pricing: $3 / $15 per M tokens

Context: 200K tokens

Where you can use it:

Claude apps, including the Free plan
Claude API, Amazon Bedrock, Vertex AI

When to reach for it:

Fast, everyday coding help
The upgrade from Claude 3.7 Sonnet

How it compares:

Level with Opus 4 on SWE-bench Verified
Behind Opus 4 on Terminal-Bench and math

What makes it different:

A fifth of Opus 4’s price
Uses tools mid-thought, like Opus 4

Read more: Introducing Claude 4 · System card

The benchmarks

Every row is Claude Opus 4, with Claude Sonnet 4 as a comparison; Anthropic reports the best of standard or extended thinking (up to 64K tokens for TAU-bench, GPQA, MMMLU, MMMU and AIME), compares against Claude 3.7 Sonnet, o3, GPT-4.1 and Gemini 2.5 Pro Preview (05-06), leaves out its parallel-compute scores here, and its Terminal-Bench figures use Claude Code as the agent (39.2% for Opus 4 and 33.5% for Sonnet 4 with the same agent as the rivals). Aider Polyglot uses 32K thinking tokens and ARC-AGI-2 uses 16K thinking for both Claude models.

Anthropic's numbers

SWE-bench Verified (Opus 4, real GitHub fixes): 72.5%. −0.2 under Claude Sonnet 4 (72.7%), +10.2 over Claude 3.7 Sonnet (62.3%), +3.4 over o3 (69.1%), +17.9 over GPT-4.1 (54.6%), +9.3 over Gemini 2.5 Pro Preview (63.2%).
Terminal-Bench (Opus 4, Claude Code agent): 43.2%. +7.7 over Claude Sonnet 4 (35.5%), +8.0 over Claude 3.7 Sonnet (35.2%), +13.0 over o3 (30.2%), +12.9 over GPT-4.1 (30.3%), +17.9 over Gemini 2.5 Pro Preview (25.3%).
GPQA Diamond (Opus 4, PhD-level science): 79.6%. +4.2 over Claude Sonnet 4 (75.4%), +1.4 over Claude 3.7 Sonnet (78.2%), −3.7 under o3 (83.3%), +13.3 over GPT-4.1 (66.3%), −3.4 under Gemini 2.5 Pro Preview (83.0%).
TAU-bench Retail (Opus 4, retail agent tasks): 81.4%. +0.9 over Claude Sonnet 4 (80.5%), +0.2 over Claude 3.7 Sonnet (81.2%), +11.0 over o3 (70.4%), +13.4 over GPT-4.1 (68.0%).
TAU-bench Airline (Opus 4, airline agent tasks): 59.6%. −0.4 under Claude Sonnet 4 (60.0%), +1.2 over Claude 3.7 Sonnet (58.4%), +7.6 over o3 (52.0%), +10.2 over GPT-4.1 (49.4%).
MMMLU (Opus 4, multilingual knowledge): 88.8%. +2.3 over Claude Sonnet 4 (86.5%), +2.9 over Claude 3.7 Sonnet (85.9%), +0.0 over o3 (88.8%), +5.1 over GPT-4.1 (83.7%).
MMMU (Opus 4, visual reasoning): 76.5%. +2.1 over Claude Sonnet 4 (74.4%), +1.5 over Claude 3.7 Sonnet (75.0%), −6.4 under o3 (82.9%), +1.7 over GPT-4.1 (74.8%), −3.1 under Gemini 2.5 Pro Preview (79.6%).
AIME 2025 (Opus 4, competition math): 75.5%. +5.0 over Claude Sonnet 4 (70.5%), +20.7 over Claude 3.7 Sonnet (54.8%), −13.4 under o3 (88.9%), −7.5 under Gemini 2.5 Pro Preview (83.0%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate, Opus 4): 20.6. +1.7 over Claude Sonnet 4 (18.9), +0.4 over o3 (20.2), +6.1 over Gemini 2.5 Pro Preview (14.5), +7.9 over GPT-4.1 (12.7).
Aider Polyglot (Opus 4, multi-language code edits): 72.0%. +10.7 over Claude Sonnet 4 (61.3%), −7.6 under o3 (high) (79.6%), +19.6 over GPT-4.1 (52.4%).
ARC-AGI-2 (Opus 4, abstract reasoning puzzles): 8.6%. +2.7 over Claude Sonnet 4 (5.9%), +2.1 over o3 (high) (6.5%).