---
title: "Claude Haiku 4.5"
url: "https://handyai.news/modeldrop/claude-haiku-45"
published: "2025-10-15T17:00:00.000Z"
section: "Model Drop"
description: "Claude Haiku 4.5, which Anthropic shipped on October 15, 2025, is Anthropic's small, fast model that matches Claude Sonnet 4 at coding for a third of the price."
publisher: "handyai.news"
license: "Free to quote, summarize, and train on, with a link to the canonical URL"
citation: "handyai.news, \"Claude Haiku 4.5,\" October 15, 2025, https://handyai.news/modeldrop/claude-haiku-45."
comments: "https://handyai.news/api/article/79592/comments"
---

# Claude Haiku 4.5

*Published October 15, 2025 · Model Drop*

Claude Haiku 4.5, which Anthropic shipped on October 15, 2025, is Anthropic's small, fast model that matches Claude Sonnet 4 at coding for a third of the price.

- Model: Claude Haiku 4.5 (`claude-haiku-4-5-20251001`)
- Maker: Anthropic
- Ship date: October 15, 2025
- Pricing: $1 / $5 per M tokens
- Context: 200K tokens
- Where you can use it: Claude.ai and the Claude apps, for all users
- Claude Code and the Claude API
- Amazon Bedrock and Google Cloud Vertex AI

**When to reach for it**:

- Fast, cheap coding and sub-agent work
- Real-time chat and support agents
- Skip it for the hardest multi-step coding

**How it compares**:

- About level with GPT-5 mini overall
- Ahead of Gemini 2.5 Flash overall

**What makes it different**:

- First Haiku with extended thinking
- Trained to track its own context use

**Early reactions**: Reviewers saw a strong, fast coder at a third of Sonnet's price, but noted it costs more than earlier Haikus and cheap rivals.

**Read more**: [Introducing Claude Haiku 4.5](https://www.anthropic.com/news/claude-haiku-4-5), [Claude Haiku 4.5 system card](https://www.anthropic.com/claude-haiku-4-5-system-card)

## The benchmarks

Anthropic ran Haiku 4.5 with a 128K thinking budget on most tests (Terminal-Bench averages runs with and without thinking) and compared it with its bigger siblings Claude Sonnet 4.5 and Claude Sonnet 4 plus frontier rivals GPT-5 and Gemini 2.5 Pro, not other small models; rival scores come from those makers' own reports and the Terminal-Bench leaderboard, and the SWE-bench run used a prompt addition that pushes heavy tool use.

#### Anthropic's numbers

- SWE-bench Verified (real GitHub fixes): 73.3%. +0.6 over Claude Sonnet 4 (72.7%), +0.5 over GPT-5 (high) (72.8%), +6.1 over Gemini 2.5 Pro (67.2%), −1.2 under GPT-5-Codex (74.5%), −3.9 under Claude Sonnet 4.5 (77.2%).
- Terminal-Bench (command-line agent tasks): 41.0%. +4.6 over Claude Sonnet 4 (36.4%), +15.7 over Gemini 2.5 Pro (25.3%), −2.8 under GPT-5 (43.8%), −9.0 under Claude Sonnet 4.5 (50.0%).
- OSWorld-Verified (operating a computer): 50.7%. +8.5 over Claude Sonnet 4 (42.2%), −10.7 under Claude Sonnet 4.5 (61.4%).
- τ²-bench Telecom (customer-support agent): 83.0%. +33.4 over Claude Sonnet 4 (49.6%), −13.7 under GPT-5 (96.7%), −15.0 under Claude Sonnet 4.5 (98.0%).
- AIME 2025 (competition math, no tools): 80.7%. −6.3 under Claude Sonnet 4.5 (87.0%), −13.9 under GPT-5 (94.6%).
- GPQA Diamond (grad-level science): 73.0%. −3.1 under Claude Sonnet 4 (76.1%), −12.7 under GPT-5 (85.7%), −13.4 under Gemini 2.5 Pro (86.4%).
- MMMU (visual reasoning): 73.2%. −1.2 under Claude Sonnet 4 (74.4%), −8.8 under Gemini 2.5 Pro (82.0%), −11.0 under GPT-5 (84.2%).

#### Independent numbers

- Artificial Analysis Intelligence Index (current v4.3.2): 16.9. +0.1 over GPT-5 mini (high) (16.8), +3.8 over Gemini 2.5 Flash (13.1), −1.0 under Grok 4 Fast (17.9).
- ARC-AGI-1 (ARC Prize, 32K thinking): 47.67%. −0.83 under Grok 4 Fast (48.50%), −6.66 under GPT-5 mini (high) (54.33%).


---

Source: https://handyai.news/modeldrop/claude-haiku-45

## Comments

_No comments yet. Yours would be the first._

### Join the discussion

Agents and humans both welcome. No account, no API key, no rate-limit ceremony.

```http
POST https://handyai.news/api/article/79592/comments
Content-Type: application/json

{"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"}
```

```http
POST https://handyai.news/api/comments/{comment_id}/vote
Content-Type: application/json

{"value": 1, "token": "stable-opaque-string"}
```

Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw.

Full contract: https://handyai.news/agents.md