Name: Claude Haiku 4.5
Author: Anthropic

Overview

Claude Haiku 4.5, released October 15, 2025, is Anthropic's entry-level tier model and the first Haiku to ship with extended thinking and computer use built in. Those two capabilities had previously been exclusive to Sonnet and Opus. Their inclusion in Haiku signals that Anthropic is moving capability downward in its lineup rather than reserving agentic features for premium SKUs.

TL;DR

First Haiku with extended thinking and computer use - previously reserved for higher tiers
73.3% on SWE-bench Verified at $1/$5 per million tokens - one-third the cost of Claude Sonnet 4.6
Beats Sonnet 4 on computer use (50.7% vs 42.2% OSWorld) while running 2x faster

At $1 per million input tokens and $5 per million output tokens, Haiku 4.5 is about one-third the price of Sonnet 4.5, with Anthropic claiming more than 2x the speed. The coding performance is real: 73.3% on SWE-bench Verified, a 500-problem real-world bug-fixing test, puts it within 4 points of Sonnet 4.5's 77.2% while costing roughly one-third as much per token. In Augment's independent agentic coding evaluation, Haiku 4.5 reached 90% of Sonnet 4.5's score - a number worth checking against because Augment runs their own tests rather than relying on vendor-supplied results.

Anthropic also made Haiku 4.5 the default model for free Claude.ai users, replacing whatever tier previously served that slot. The company framed it as democratizing near-frontier intelligence. From a product standpoint, it's also a competitive move against OpenAI, which serves GPT-4o mini on its free tier.

Key Specifications

Specification	Details
Provider	Anthropic
Model Family	Claude
Parameters	Not disclosed
Context Window	200K tokens
Max Output	64K tokens
Input Price	$1.00/M tokens
Output Price	$5.00/M tokens
Cache Read Price	$0.10/M tokens
Release Date	October 15, 2025
Knowledge Cutoff	February 2025
License	Proprietary
Safety Level	ASL-2

Benchmark Performance

Claude Haiku 4.5's headline number is SWE-bench Verified, which tests models on unmodified GitHub issues. The 73.3% score is averaged across 50 runs on the full 500-problem dataset with no test-time compute tricks - the methodology Anthropic disclosed in its system card. That score beats the previous-generation Sonnet 4 (72.7%) outright.

Benchmark	Claude Haiku 4.5	Claude Sonnet 4.5	Sonnet 4
SWE-bench Verified	73.3%	77.2%	72.7%
OSWorld (computer use)	50.7%	56.0%	42.2%
MMMU (visual reasoning)	73.2%	74.4%	~72%
Output speed	~93 t/s	~45 t/s	~40 t/s

SWE-bench: averaged over 50 runs, 128K thinking budget, no test-time compute. OSWorld: GUI control benchmark. Speed: Artificial Analysis independent measurement.

The computer use score is the more surprising result. OSWorld tests a model's ability to control a desktop GUI - click buttons, fill forms, navigate interfaces - and Haiku 4.5 scores 50.7% versus Sonnet 4's 42.2%. Haiku now outperforms a previous generation Sonnet on autonomous computer control, at less than half the price.

On MMMU, a visual reasoning benchmark covering images, charts, and technical diagrams, Haiku 4.5 scores 73.2% versus Sonnet 4's 74.4%. The gap is about 1 percentage point. For workloads that require parsing screenshots or documents, the price-to-performance tradeoff tilts toward Haiku.

Speed is a structural advantage. Artificial Analysis measured 93.1 tokens per second - roughly double the ~45 t/s pace of Sonnet-tier models. For real-time applications where latency matters more than output quality, that speed difference is often the deciding factor.

Abstract neural network visualization representing AI computation Claude Haiku 4.5 uses a dense transformer architecture with absolute position embeddings and context-awareness training to manage long agentic workflows. Source: unsplash.com

Key Capabilities

Extended thinking. Haiku 4.5 is the first model in the Haiku line to support chain-of-thought reasoning with a configurable thinking budget. You set a token budget up to 128K; the model reasons internally before responding. For multi-step coding or debugging tasks, this lets a cheap model stretch into territory that previously required a Sonnet or Opus tier call.

Computer use. The model can interpret screenshots and issue mouse and keyboard actions - cursor movement, button clicks, form fills, and application navigation. Anthropic's OSWorld score (50.7%) is the external validation here. The 2023-vintage Sonnet 3.5 scored 14% on the same benchmark; Haiku 4.5 scores 50.7%. That's not a minor version bump.

Context awareness. Anthropic built explicit context-awareness training into Haiku 4.5 to address what they call "agentic laziness" - the tendency of models to quit early or skip steps in long workflows. The model tracks its remaining token budget after each tool call and adjusts pacing accordingly: winding down when context runs low, maintaining momentum when headroom remains. In agentic pipelines processing hundreds of tool calls, this matters.

Multi-agent sub-agent role. The canonical deployment pattern pairs Sonnet 4.5 as orchestrator with multiple Haiku 4.5 instances executing parallel subtasks. A Sonnet breaks a problem into a plan; Haiku workers execute steps simultaneously. The economics make this viable: running 10 Haiku calls costs the same as roughly 3 Sonnet calls. GitLab's Duo Agentic Chat runs exactly this pattern.

Haiku 4.5 supports both text and image inputs, with text-only output. The 200K context window is standard across Anthropic's current lineup. Maximum output is 64K tokens - up from 8,192 in Haiku 3.5, which matters for code generation tasks that require long file outputs.

Pricing and Availability

At $1/$5 per million tokens, Haiku 4.5 fits between ultra-budget competitors and capable mid-tier models. GPT-4o mini costs $0.15/$0.60 - dramatically cheaper per token. Gemini 2.0 Flash Lite sits at $0.075/$0.30. For pure volume workloads on commodity tasks, those models undercut Haiku 4.5 significantly.

The counter-argument is task quality. Haiku 4.5's 73.3% SWE-bench score isn't matched by GPT-4o mini or Gemini Flash variants in that benchmark tier. If your workload is coding-heavy or agentic rather than commodity text generation, the per-task economics may still favor Haiku 4.5 despite the higher per-token cost.

Prompt caching is available at $0.10 per million tokens for cache reads (a 92% discount on input price), which changes the math significantly for agentic workflows with repeated system prompts or tool definitions. A pipeline that front-loads a 50K-token system prompt and calls it 1,000 times costs roughly $0.10 in cache reads versus $50 in uncached input tokens.

The model is available on the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI. All free Claude.ai users get Haiku 4.5 access by default. Paid Claude Pro and Team subscribers get higher rate limits and priority access to Sonnet and Opus tiers.

Anthropic's system card for Haiku 4.5 classifies it at ASL-2, compared to ASL-3 for Sonnet 4.5 and Opus 4.1. The lower classification reflects Haiku's assessed risk profile on biosecurity and other high-stakes domains - not a downgrade on alignment, where the system card notes Haiku 4.5 shows fewer misaligned behaviors than both Sonnet 4.5 and Opus 4.1 in head-to-head evaluations.

Robot figure representing AI automation capabilities Computer use in Haiku 4.5 enables direct GUI control, with a 50.7% success rate on OSWorld - higher than the previous-generation Sonnet 4 model. Source: unsplash.com

Strengths

73.3% on SWE-bench Verified at $1/$5/M - the best coding-to-price ratio in Anthropic's lineup
93 tokens/second output speed, roughly 2x faster than Sonnet-tier models
First Haiku with extended thinking (128K budget) and computer use
50.7% on OSWorld - beats previous-generation Sonnet 4 on GUI control
Context-awareness training reduces agentic task abandonment
ASL-2 with lower misaligned behavior rates than larger Anthropic models
Prompt caching at $0.10/M cache reads cuts costs 92% on repeated contexts

Weaknesses

Priced above GPT-4o mini ($0.15/$0.60) and Gemini 2.0 Flash Lite ($0.075/$0.30) on a per-token basis
Knowledge cutoff of February 2025 - same as other Claude 4.x models
No native audio or video input; text and images only
Parameters not disclosed, so external efficiency comparisons are limited
MMMU visual reasoning (73.2%) trails GPT-5 variants (84.2%) by a significant margin
Context window tops at 200K - Gemini 2.5 Flash offers 1M tokens for long-document work

Claude Sonnet 4.6 model card - mid-tier comparison
Claude Opus 4.6 model card - flagship comparison
Computer Use Leaderboard - full OSWorld rankings
Coding Benchmarks Leaderboard - SWE-bench context
Agentic AI Benchmarks Leaderboard - multi-agent performance

FAQ

Is Claude Haiku 4.5 free to use?

Yes. Anthropic made Haiku 4.5 the default model for all free Claude.ai users. Rate limits apply; paid plans get higher throughput.

Does Claude Haiku 4.5 support extended thinking?

Yes. It's the first Haiku model to do so. The thinking budget goes up to 128K tokens and is configurable per API call.

How does Haiku 4.5 compare to Claude Sonnet 4.5 on coding?

Haiku 4.5 scores 73.3% on SWE-bench Verified; Sonnet 4.5 scores 77.2%. The 4-point gap costs roughly 3x less per token to run.

What is context awareness in Claude Haiku 4.5?

A training technique that teaches the model to track its remaining token budget during long agentic workflows, preventing premature task abandonment.

Is Claude Haiku 4.5 available on Amazon Bedrock?

Yes. Haiku 4.5 is available on Anthropic's API, Amazon Bedrock, and Google Cloud Vertex AI.

Sources: