Claude Haiku 4.5
Anthropic's small, fast Claude model for coding, agents and real-time work
Claude Haiku 4.5 is a small language model from Anthropic, released in October 2025 as the newest member of the Claude family. The company positions it as a fast, low-cost option that reaches coding performance close to Claude Sonnet 4 while running more than twice as quickly at a third of the cost.
- Coding performance close to Claude Sonnet 4 at a third of the cost
- Runs more than twice as fast as Sonnet 4, Anthropic says
- 73.3% on SWE-bench Verified, averaged over 50 trials
- Released under AI Safety Level 2, below the ASL-3 used for Sonnet 4.5
- API pricing of $1 per million input and $5 per million output tokens
What it is
Claude Haiku 4.5 is a small general-purpose language model made by Anthropic and released in October 2025. It is the newest entry in the Claude line and is aimed at work where speed and price matter as much as quality: chat assistants, customer service agents, pair programming and software agents that act on a computer. Anthropic says it reaches coding performance close to Claude Sonnet 4, a model that sat at the frontier five months earlier, while costing a third as much and running more than twice as fast. On some tasks, such as controlling a computer, the company says it beats Sonnet 4.
How it came about
Anthropic released Claude Haiku 4.5 to all users at once, two weeks after putting out Claude Sonnet 4.5, which it still describes as its frontier model. The company frames the release as moving recent frontier capability into a cheaper, faster tier. It replaces Claude Haiku 3.5 and can also stand in for Sonnet 4. Anthropic suggests pairing the two new models: Sonnet 4.5 splits a hard problem into steps, then directs several Haiku 4.5 instances working on subtasks in parallel.
Inside the model
Anthropic has not published layer counts, parameter totals or training data for Claude Haiku 4.5 in this release. Evaluation runs used extended thinking with a 128K thinking budget and default sampling settings, and a smaller 32K budget in some Terminal-Bench runs. The model supports tool use and computer control, and was tested with scaffolds that give it bash and file editing. Anthropic reports it was aligned and fine-tuned to reduce misaligned behaviour, and says automated assessment found fewer such behaviours than in Sonnet 4.5 or Opus 4.1.
What it can do
Anthropic reports 73.3% on SWE-bench Verified, averaged over 50 trials on the full 500-problem set. Terminal-Bench scores were 40.21% without thinking and 41.75% with a 32K thinking budget. Other reported evaluations include tau2-bench for tool-using agents, AIME for maths, OSWorld for computer control and MMMLU across 14 non-English languages. The company says it matches Sonnet 4 on coding while running more than twice as fast, and that it beats Sonnet 4 on computer use. Partner firms quoted by Anthropic describe it as reaching 90% of Sonnet 4.5's score in one agentic coding test and running four to five times faster.
How to use it
The model is sold through the Claude API under the identifier claude-haiku-4-5, priced at $1 per million input tokens and $5 per million output tokens. It is also on Amazon Bedrock and Google Cloud's Vertex AI, and inside Anthropic's own apps and Claude Code. Anthropic describes it as a drop-in replacement for Haiku 3.5 and Sonnet 4 at its cheapest price point. Weights are not released.
Safety and testing
Anthropic says it ran a series of safety and alignment evaluations and saw low rates of concerning behaviour, with the model substantially more aligned than Claude Haiku 3.5. Its automated assessment recorded a statistically significantly lower overall rate of misaligned behaviour than Sonnet 4.5 and Opus 4.1, which the company calls its safest model on that measure. Testing found only limited risk from chemical, biological, radiological and nuclear weapons work, so the model ships under the AI Safety Level 2 standard, while Sonnet 4.5 and Opus 4.1 use the stricter ASL-3. A system card gives the full reasoning.
Reception and debate
Firms given early access, including Augment, Warp and GitHub, describe the model as fast and cost-efficient for coding and agent work. One partner reported 65% accuracy on instruction-following for slide text generation against 44% from its own premium tier. These are vendor comments collected by Anthropic rather than independent reviews. No lawsuits, bans or public incidents are reported here.
Why it matters
Anthropic presents Haiku 4.5 as showing that near-frontier coding quality can now sit in a cheap, fast tier, and as a way to run many model instances together under a stronger planner. The company points to Claude for Chrome and Claude Code as products made faster by it, and partners describe it as opening up real-time and multi-agent uses that were too slow or costly before.
Questions people ask about Claude Haiku 4.5
What is Claude Haiku 4.5 for?
Anthropic built it for work where replies must come quickly and costs stay low: chat assistants, customer service, pair programming and software agents. The company says it reaches coding quality near Claude Sonnet 4 while running more than twice as fast at a third of the price.
How do I get it?
Developers call it through the Claude API as claude-haiku-4-5, or through Amazon Bedrock and Google Cloud's Vertex AI. It is also inside Anthropic's apps and Claude Code. Anthropic prices it at $1 per million input tokens and $5 per million output tokens.
How large is the model?
Anthropic has not given a parameter count, layer count or training data size for Claude Haiku 4.5 in this release. It describes the model only as its latest small model, aimed at speed and cost efficiency.
How does it compare with other Claude models?
Anthropic keeps Claude Sonnet 4.5 as its frontier model and best coding model. Haiku 4.5 is meant to sit just below that level at much lower cost, and the company says it beats Sonnet 4 on tasks such as computer use. It replaces Claude Haiku 3.5.
What safety level does it carry?
Anthropic released it under AI Safety Level 2 after testing found only limited risk from chemical, biological, radiological and nuclear weapons work. Sonnet 4.5 and Opus 4.1 use the stricter ASL-3 standard. The company says its automated assessment found fewer misaligned behaviours in Haiku 4.5 than in either of those models.
Good reading and sources
There is no English Wikipedia article on Claude Haiku 4.5 yet, so the facts on this page come from the sources listed above: anthropic.com/news/claude-haiku-4-5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.