Model Drop

Claude Sonnet 4.5

September 29, 2025 2 min read

Claude Sonnet 4.5, which Anthropic shipped on September 29, 2025, is a mid-priced Claude model that Anthropic pitched as the world's best for coding, building agents, and using computers.

Model: Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
Maker: Anthropic
Ship date: September 29, 2025
Pricing: $3 / $15 per M tokens
Context: 200K tokens
Where you can use it:
- Claude apps, with code execution on paid
- Claude Code, now with checkpoints
- Claude API and cloud partners
When to reach for it:
- Long, multi-step autonomous coding runs
- Browser and computer-use agents
- Skip it if lowest token price matters most
How it compares:
- Ahead of GPT-5 and Opus 4.1 on agent coding
- Behind GPT-5 and Gemini 2.5 Pro on hard math
What makes it different:
- Ships with the Claude Agent SDK
- Context editing and a memory tool in API
Early reactions: Early testers called it the best coding model for now, ahead of GPT-5-Codex in hands-on use.
Read more: Introducing Claude Sonnet 4.5, System card

The benchmarks

Anthropic compares Sonnet 4.5 with its own Opus 4.1 and Sonnet 4 plus GPT-5 and Gemini 2.5 Pro, using rival numbers from their makers and public leaderboards; SWE-bench Verified uses a 200K thinking budget with no parallel compute (82.0% with it), Terminal-Bench uses the Terminus 2 agent for Claude and GPT-5 but Terminus 1 for Gemini, τ²-bench uses prompt addenda, OSWorld scores were run on the OSWorld-Verified framework, and AIME is shown without tools (100% with Python). ARC-AGI-2 uses 32K thinking for Claude and Gemini, and the LMArena row is the October 1, 2025 snapshot (2,600 votes, standard mode).

Anthropic's numbers

SWE-bench Verified (real GitHub fixes): 77.2%. +2.7 over Claude Opus 4.1 (74.5%), +4.5 over Claude Sonnet 4 (72.7%), +4.4 over GPT-5 (72.8%), +2.7 over GPT-5-Codex (74.5%), +10.0 over Gemini 2.5 Pro (67.2%).
Terminal-Bench (terminal coding tasks): 50.0%. +3.5 over Claude Opus 4.1 (46.5%), +13.6 over Claude Sonnet 4 (36.4%), +6.2 over GPT-5 (43.8%), +24.7 over Gemini 2.5 Pro (25.3%).
OSWorld (computer use tasks): 61.4%. +17.0 over Claude Opus 4.1 (44.4%), +19.2 over Claude Sonnet 4 (42.2%).
τ²-bench Retail (retail support agent): 86.2%. −0.6 under Claude Opus 4.1 (86.8%), +2.4 over Claude Sonnet 4 (83.8%), +5.1 over GPT-5 (81.1%).
τ²-bench Telecom (telecom support agent): 98.0%. +26.5 over Claude Opus 4.1 (71.5%), +48.4 over Claude Sonnet 4 (49.6%), +1.3 over GPT-5 (96.7%).
AIME 2025 (competition math): 87.0%. +9.0 over Claude Opus 4.1 (78.0%), +16.5 over Claude Sonnet 4 (70.5%), −7.6 under GPT-5 (94.6%), −1.0 under Gemini 2.5 Pro (88.0%).
GPQA Diamond (PhD-level science): 83.4%. +2.4 over Claude Opus 4.1 (81.0%), +7.3 over Claude Sonnet 4 (76.1%), −2.3 under GPT-5 (85.7%), −3.0 under Gemini 2.5 Pro (86.4%).
MMMU (visual reasoning): 77.8%. +0.7 over Claude Opus 4.1 (77.1%), +3.4 over Claude Sonnet 4 (74.4%), −6.4 under GPT-5 (84.2%), −4.2 under Gemini 2.5 Pro (82.0%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 20.7. −2.3 under GPT-5 (23.0), +4.6 over Gemini 2.5 Pro (16.1), −1.8 under Grok 4 (22.5), −2.1 under Claude Opus 4.1 (22.8).
ARC-AGI-2 (abstract reasoning puzzles): 13.6%. +3.7 over GPT-5 (high) (9.9%), +8.7 over Gemini 2.5 Pro (4.9%), −2.4 under Grok 4 (16.0%).
LMArena text Elo (human preference votes): 1419. −47 under Gemini 2.5 Pro (1466), −1 under Claude Opus 4.1 (thinking) (1420), +9 over GPT-5 (high) (1410), −3 under Grok 4 (1422).