Haiku 4.5: Sonnet 4 performance, 3x cheaper (now available in Kilo Code)
The AI coding model race heats up.
We’re excited to announce that Kilo Code now supports Claude Haiku 4.5, Anthropic’s latest small model that delivers frontier-level coding performance at a fraction of the cost.
What Makes Haiku 4.5 Special
Five months ago, Claude Sonnet 4 was state-of-the-art model. Today, Haiku 4.5 matches its coding performance while running faster and costing 3x+ less.
Is Anthropic going after xAI and Grok Code Fast 1 with this release? Let’s check some interesting benchmark stats.
Third-party benchmark stats:
73.3% on SWE-bench Verified (averaged over 50 trials)
41% on Terminal-Bench for command-line operations
Comparable to Sonnet 4 (according to Anthropic’s announcment)
90% of Sonnet 4.5’s performance in Augment’s agentic coding evaluation
Perfect for Real-Time Development
According to Anthropic, Haiku 4.5 is built for speed without sacrificing intelligence. It’s ideal for:
Pair programming with near-instant responses
Rapid prototyping with minimal latency
Real-time code completion and suggestions
Command-line operations and tool calling (using an AI coding agent like Kilo Code)
Haiku 4.5 has a 200K token context window, which makes it comparable to other AI models of its class (like Grok Code Fast 1).
How to use it with Sonnet 4.5
According to Anthropic, Sonnet 4.5 still remains “the world’s best coding model.”
However, Haiku 4.5 opens up a new set of workflows. You can use these 2 models together:
Sonnet 4.5 to break down complex problems and create multi-step plans
Haiku 4.5 to execute subtasks in parallel
This way, you’ll get frontier-level results at dramatically lower cost and latency.
Available in Kilo Code (top-up $10, get $20 extra)
Haiku 4.5 is available now in Kilo Code (just pick claude-haiku-4-5 from your list of models):
Our extension is 100% free & open source. We also have a gateway through which you can use Haiku (instead of buying via external API providers).
If you haven’t topped up, you’ll get $20 extra if you top-up $10 or more:
That’s 30 million input tokens for Haiku 4.5 (when working with AI coding agents like Kilo Code, 70%+ of your token costs are input token costs).
Kilo Code is available as an extension for VS Code and Jet Brains (we support all JetBrains IDEs, such as PyCharm, WebStorm, IntelliJ idea, and so on).
These updates are nice but y’all gotta stop writing it with (I’m assuming) Claude lol. Little human touch goes a long way
“This isn’t just an incremental update—it’s a fundamental shift in what’s possible with smaller models.”
The combination of the common AI-phraseology ‘This isn’t just __, it’s a ____’ along with the use of the em-dash led me to believe that there was AI-generated writing here
There have also been other posts on the Kilo blog that have the familiar AI ‘tourist website tone’
Apologies if my conclusion was unfounded or a bit gung-ho, I just enjoy the tool and was peeved not at a perception of AI-enhanced-writing (nothing wrong with that), just at perceived slop on the blog of a tool I enjoy