Claude Haiku 4.5 Release: Fastest Claude Model Gets Quality Boost with 73.3% SWE-Bench Score

A $1-per-million-token model just matched the coding quality of a model that costs three times more. Claude Haiku 4.5, released on October 15, 2025, isn’t just another incremental update — it’s Anthropic’s clearest signal yet that frontier-level AI doesn’t have to come with frontier-level pricing.

When Anthropic first launched the Haiku line, it was positioned as the “fast and cheap” option — good for simple tasks, but you’d reach for Sonnet or Opus when things got serious. That calculus just changed dramatically. Claude Haiku 4.5 scores 73.3% on SWE-bench Verified, putting it neck-and-neck with Claude Sonnet 4, which was the state-of-the-art coding model just months ago. One developer on the InfoQ forums put it bluntly: “I’ve never built apps so fast, and it does super well. I don’t even need Claude Sonnet anymore.”

Claude Haiku 4.5 Benchmark Performance: The Numbers That Matter

Let’s cut straight to the benchmarks, because that’s where Claude Haiku 4.5 tells its most compelling story. The headline number is 73.3% on SWE-bench Verified — a test that measures a model’s ability to solve real-world software engineering tasks from actual GitHub issues. For context, Claude Sonnet 4 achieved a similar score when it was considered the gold standard for AI-assisted coding in spring 2025.

But SWE-bench is just one piece of the puzzle. Claude Haiku 4.5 also posted a 41.75% score on Terminal-Bench with extended thinking enabled, showing strong performance on complex terminal-based coding tasks. On the AIME mathematical reasoning benchmark, the model demonstrated high-level problem-solving capabilities, and on OSWorld — which tests computer interaction tasks — Claude Haiku 4.5 actually exceeds Claude Sonnet 4’s performance.

Perhaps the most telling metric comes from Anthropic’s own agentic coding evaluation: Claude Haiku 4.5 achieves 90% of Sonnet 4.5’s performance. That’s the current flagship model being closely matched by the budget tier. Epoch AI researchers found that “even with reasoning disabled, Haiku 4.5 performs similarly or better than early lightweight reasoning models, like o1-mini.”

Extended Thinking and Computer Use: Premium Features Go Budget

The most significant architectural addition to Claude Haiku 4.5 is extended thinking — a feature previously reserved for the more expensive Sonnet and Opus models. This gives Haiku a 128K thinking budget for complex reasoning tasks, essentially allowing the model to “think longer” before responding. In practice, this means Claude Haiku 4.5 now operates in two distinct modes:

• Standard mode: Rapid responses for straightforward queries — the classic Haiku speed advantage at up to 4-5x faster than Sonnet 4.5
• Extended thinking mode: Allocates additional processing time for multi-step reasoning, complex code generation, and analytical tasks where accuracy matters more than speed

Computer use is the other headline feature making its Haiku debut. This capability allows Claude Haiku 4.5 to interact with computer interfaces — clicking buttons, filling forms, navigating applications — which opens the door for agentic workflows at a fraction of the cost. And it’s not just “good enough” — on the OSWorld benchmark for computer interaction tasks, Haiku 4.5 actually outperforms Sonnet 4.

Pricing: One-Third the Cost of Sonnet 4

Claude Haiku 4.5 is priced at $1 per million input tokens and $5 per million output tokens. To put that in perspective: Claude Sonnet 4.5 costs $3/$15, and Claude Opus runs at $15/$75. For teams running high-volume AI workloads — think automated code review, customer support, or data extraction pipelines — this pricing changes the unit economics completely.

The model supports a 200,000-token context window and up to 64,000 tokens of output, matching the context capabilities of the more expensive Sonnet models. For most production use cases, you’re getting 90% of the capability at roughly 33% of the cost — a ratio that makes Claude Haiku 4.5 the default choice for any cost-sensitive deployment.

Same-Day GitHub Copilot Integration

In a notable move, GitHub announced same-day availability of Claude Haiku 4.5 in Copilot. The model is accessible across all Copilot tiers — Pro, Pro+, Business, and Enterprise — and works in VS Code’s chat, ask, edit, and agent modes. This is the first time a Haiku-class model has been integrated into GitHub Copilot, reflecting the quality leap this release represents.

For developers already using Copilot, this means access to near-Sonnet coding quality at Haiku speeds directly in their IDE. The recommended VS Code version is 1.105 or higher for optimal performance. Enterprise and Business users will need their administrators to enable the Claude Haiku 4.5 policy in Copilot settings first.

Safety: Anthropic’s Most Aligned Model Yet

According to Anthropic’s system card, Claude Haiku 4.5 carries an ASL-2 safety classification — less restrictive than Sonnet 4.5’s ASL-3, but with measurably better alignment. The model shows “substantially more alignment” than its predecessor Claude Haiku 3.5, and Anthropic’s automated testing found it has a statistically significantly lower rate of misaligned behaviors than both Claude Sonnet 4.5 and Claude Opus 4.1.

This makes Claude Haiku 4.5 Anthropic’s safest model by their internal metrics — an unusual achievement for a budget-tier model. It suggests that Anthropic’s alignment techniques are scaling efficiently, delivering better safety outcomes without requiring the computational overhead of larger models.

Multi-Agent and Production Use Cases

Where Claude Haiku 4.5 really shines is in multi-agent architectures. When you’re orchestrating dozens or hundreds of AI agents — each one handling a subtask like code review, data validation, or content processing — the cost per agent matters enormously. At $1/$5 per million tokens, you can run five Haiku agents for the price of one Sonnet agent, with 90% of the capability.

The model is available across all major platforms from day one: Anthropic’s API, Amazon Bedrock, and Google Cloud’s Vertex AI. This broad availability, combined with the pricing and performance profile, positions Claude Haiku 4.5 as the go-to model for production AI systems where cost-efficiency and reliability are non-negotiable.

The Bigger Picture: Budget Models Are Catching Up Fast

Claude Haiku 4.5’s release arrives during a significant week in tech — Meta Connect and Adobe MAX are both happening, flooding the AI conversation with announcements. But Anthropic’s move might be the most consequential for developers. The gap between “budget” and “premium” AI models is collapsing faster than anyone predicted. A model that costs $1 per million input tokens now matches what was state-of-the-art six months ago.

For teams building AI-powered products, the strategic implication is clear: you can likely downgrade your model tier without sacrificing meaningful quality. The 73.3% SWE-bench score, extended thinking capability, and computer use support make Claude Haiku 4.5 not just a budget option — it’s a genuinely capable model that happens to be priced aggressively. The question isn’t whether to try it. It’s whether you can still justify paying three times more for Sonnet.