Oct 15, 2025

Introducing Claude Haiku 4.5: Faster, Cheaper, Near-Frontier AI Coding

Claude Haiku 4.5 is available now. Frontload the conclusion: it delivers near-frontier coding quality while cutting cost and latency substantially. If your work depends on low latency and predictable token economics then this model changes the practical defaults for how you compose agentic systems and copilots.

What changed with Haiku 4.5

The simple way to think about it is this. Haiku 4.5 closes the gap between small and frontier models on a set of real engineering tasks while running faster and cheaper. Compared to Sonnet 4 which was state of the art a few months ago, Haiku 4.5 achieves similar coding outcomes at roughly one third the cost and more than twice the runtime speed in many workflows. Partners report it runs four to five times faster than Sonnet 4.5 on some real-time agent loops.

That cost and latency profile matters more than raw single-shot accuracy for many production systems. If you build IDE copilots, customer service agents, or multi-agent coding pipelines, Haiku 4.5 makes previously expensive designs practical.

How it changes architectures

There is a clear pattern to adopt. Use Sonnet 4.5 as the planner for difficult, multi-step reasoning and use Haiku 4.5 as the executor for high-throughput, low-latency subtasks. That pattern converts expensive monolithic agents into a two tier system where the planner escalates only the hardest parts. This reduces cost, speeds up wall-clock time, and keeps the heavier model reserved for tasks that truly need it. If you want more on router tradeoffs and why picking the right model per step matters read my piece on model routers and control here: https://adam.holter.com/model-routers-for-llms-reliability-wins-quality-suffers-without-control/

Guy Gur-Ari Co-Founder of Augment Code: “Claude Haiku 4.5 hit a sweet spot we didn’t think was possible: near-frontier coding quality with blazing speed and cost efficiency. In Augment’s agentic coding evaluation, it achieves 90% of Sonnet 4.5’s performance, matching much larger models. We’re excited to offer it to our users.”

Benchmarks that matter for buying decisions

Benchmarks are noisy but some metrics are directly relevant to purchasing choices.

SWE-bench Verified reached 73.3 percent averaged over 50 trials using a two-tool scaffold with bash and file editing and default sampling settings. That performance was achieved with a prompt addendum encouraging frequent tool use and writing tests first.
Partner tests show Haiku 4.5 matches or exceeds Sonnet 4 on computer use tasks that mimic real tool chains, which helps explain why it feels quicker in interactive UIs.
Instruction-following for slide text generation scored 65 percent accuracy in partner evaluation versus 44 percent from a previous premium tier model. That directly affects unit economics for large scale document pipelines.

The above chart is based off of vibes only. Source: Trust me, Bro.

Pricing and where to run it

Pricing is straightforward. Haiku 4.5 is priced at 1 dollar per million input tokens and 5 dollars per million output tokens. It is available through the Claude API under the model name claude-haiku-4-5 and on managed platforms including Amazon Bedrock and Google Cloud Vertex AI. GitHub Copilot has Haiku 4.5 in public preview for Pro, Pro Plus, Business, and Enterprise which makes it an easy choice for teams that prioritize snappy editor experiences.

Safety and alignment notes

Anthropic ran a broad set of automated and manual tests. Haiku 4.5 showed lower rates of misaligned behaviors than Sonnet 4.5 and Opus 4.1 in the lab tests reported. That metric made it their safest model by those automated measures. Because it poses limited risk in high consequence CBRN domains it was released at AI Safety Level 2 which permits broader access than the more restrictive level assigned to some larger models. For the full details and system card see the official system card page here: https://www.anthropic.com/claude-haiku-4-5-system-card

Where Haiku 4.5 fits in real builds

I see three clear immediate use patterns.

Low-latency copilots in editors and web assistants where sub-second response is required. Haiku 4.5 reduces perceived friction and keeps token costs predictable.
Claude Code: This model is basically made for Claude code and making Claude code faster. Use Sonnet 4.5 to break down problems and use Haiku 4.5 to execute and test in parallel. That reduces end-to-end turnaround without sacrificing final quality.
Content assembly where instruction following and throughput matter. Document and slide generation pipelines benefit from the accuracy per dollar improvement.

Method notes that affect results

If you are reproducing results pay attention to these details because they matter as much as the model name.

The SWE-bench Verified score used a simple two-tool scaffold with bash and file editing and included a prompt instruction to prefer heavy tool use and write tests first.
Agent benchmarks used fixed agent frameworks, step limits, and large thinking budgets in some runs. That reduces variance but shifts the cost profile in practice.
Language and instruction-following tests averaged multiple runs across non-English languages which lowers noise but may not reflect performance on a specific dialect or phrasing.

Caveats and common operational mistakes

Near-frontier performance does not mean use Haiku 4.5 everywhere. For deep, long-horizon reasoning or single-shot tasks that need the maximum chain of thought keep Sonnet 4.5 in the loop.
Tool heavy agents still incur token costs. Haiku 4.5 is cheaper but repeated tool calls add up.
Computer use improvements are useful but not a replacement for robust guardrails. Add timeouts, watchdogs on tool call counts, and backoff retry policies so UI automation does not loop indefinitely.

How to try it in minutes

Pick your platform. Use the Claude API with model name claude-haiku-4-5 or try it on OpenRouter.
Start with a minimal coding agent scaffold. Two tools are enough: bash and file edit. Add one instruction that requires writing tests first.
Set strict budgets for tool calls and tokens and add one backoff retry policy for transient tool failures.
Run a small evaluation on a representative ticket or repo and measure pass rate, time to first useful output, and cost per solved issue. Those three numbers are the practical signal you should use to decide whether to adopt Haiku 4.5.

Final practical take

This release is not trying to be everything. It is designed to be the right default for execution where latency and unit economics matter. If your constraints match that profile, Haiku 4.5 should be your go to executor and Sonnet 4.5 should be the planner you call when the reasoning load is high. That split makes multi-agent systems and interactive copilots far more practical to run at scale.