Haiku 4.5 - Small Beats Big

Claude Haiku 4.5 is arriving with higher prices, but it’s also delivering a rare mix of speed and task performance that makes it a strong candidate to become the default “grunt work” model for agent building—especially when paired with larger models for planning and reasoning.

Pricing is the immediate sticking point. Claude Haiku 4.5 is listed at $1 per million input tokens and $5 per million output tokens. That’s a step up from earlier Haiku tiers, where Haiku 3.5 was $4 per million output tokens and Haiku 3 was $1.25 per million output tokens. On cost alone, it would be easy to dismiss the upgrade—particularly for teams that previously relied on Haiku for cheap, high-throughput workloads. But the performance comparisons suggest the higher output price may be offset by faster execution and better results on agent-style tasks.

Anthropic positions Haiku 4.5 as a model that can outperform Claude Sonnet 4 on several benchmarks, including Swebench and other agent benchmarks, plus “computer use” style evaluations. The reasoning gap isn’t framed as a full replacement for the top-tier models; instead, Haiku 4.5 is portrayed as a fast, capable workhorse. The practical workflow implied is straightforward: use a frontier model (Sonnet 4.5 and potentially Opus 4.5) for planning and deeper reasoning, then delegate execution-heavy steps—tool use, structured outputs, coding tasks, and other repetitive agent actions—to Haiku 4.5.

The transcript emphasizes that Haiku 4.5 is not just competitive on intelligence metrics; it’s also faster. Benchmarks are described as showing Haiku 4.5 beating models such as GPT-5 and Gemini 2.5 Pro on coding and other tasks, while also running about twice as fast as Sonnet 4. That speed matters because agent systems increasingly depend on rapid iterations: calling tools, producing structured outputs, and completing multi-step tasks where latency compounds.

A cost comparison further supports the “workhorse” framing. Haiku 4.5 is described as costing about one-third of what Sonnet 4.5 will cost, making it attractive as the model that handles the bulk of agent execution even if Sonnet remains the best coding and frontier option. In other words, the upgrade isn’t about replacing the smartest model—it’s about making the delegated layer cheaper per completed task.

To ground the claims, the transcript reports timing tests run through GCP, using measures like time to first token and total response time. Claude Haiku 4.5 is reported to return a response in about 3.6 seconds, with roughly half a second to the first token. Sonnet 4.5 shows longer time-to-first-token and longer full response time, while Sonnet 4 is slower still on first-token timing. Compared with earlier Haiku versions, Haiku 4.5 shows a meaningful speed bump over Haiku 3.5, and it’s described as balancing intelligence improvements with faster or comparable generation times versus Claude 3 Haiku.

Overall, the transcript paints Haiku 4.5 as a practical default for agent “grunt work”: fast tool-using execution, structured classification, and high-volume coding or task steps—while reserving the heaviest reasoning for larger models.