Claude Haiku: Anthropic's Fast, Low-Cost Line
Current status
The current flagship of the line is Claude Haiku 4.5 (released 15 October 2025). Unlike Sonnet, Opus and Fable, Haiku isn’t on generation 5 yet — details on price, context window and benchmarks live in the glossary entry.
What Claude Haiku stands for
Haiku is the bottom of the three classic Claude lines — and the only one where speed and price are the first selection criterion, not the second. Since Claude 3, Anthropic has split its models into three sizes: Opus for the hardest work, Sonnet as the balanced middle, Haiku for everything that has to happen fast and in large numbers. What’s striking is how far the bar has moved: on many tasks Haiku 4.5 reaches the level that marked the Sonnet class a year earlier — at a fraction of the cost.
One thing to keep straight: Haiku deliberately lags in the version cycle. While Sonnet, Opus and the new Fable tier are already on generation 5, Haiku 4.5 from October 2025 remains the current model of the line as of September 2026.
- Coding 3.5 / 5 · Claude Opus 5 u. a.
- Reasoning 3 / 5 · Claude Opus 5 u. a.
- Text 4 / 5 · Claude Fable 5.1
- Vision 3.5 / 5 · Gemini 3.1 Pro
- Speed 5 / 5 · Feld-Spitze
- Kosten-Eff. 4.5 / 5 · GPT Luna u. a.
Eignung 0–5 · redaktionelle Einordnung, kein Benchmark · gestrichelt = Feld-Bestwert je Achse
The profile is the mirror image of Opus: top marks on speed and cost efficiency, solid on text and coding, deliberately restrained on deep reasoning. That’s exactly what the line is built for.
Where it fits
Haiku earns its keep on tasks that demand high volume and low latency:
- Classification and extraction — routing tickets, pulling entities from documents, filling structured fields, thousands of times a day.
- Real-time chat support — short response times are a feature here, not just nice to have.
- Fast parallel sub-agent — under an orchestrating Sonnet or Opus, working through the many small sub-steps at the same time. This pattern is exactly what makes Haiku valuable in multi-agent systems.
Gotcha: no effort parameter
No model in the Haiku line supports the effort levels Anthropic offers on the larger models. Anthropic’s model spec explicitly lists “Default effort: Not supported” for Haiku 4.5. You steer reasoning depth instead through a manual, numeric thinking budget (budget_tokens). If you carry code over from a larger model, you have to adjust this spot — otherwise the call fails.
How to choose: Haiku, Sonnet or Opus?
The three lines solve the same underlying problem at different points on the curve of capability, speed and price. The positioning map shows where Haiku stands in the field:
Redaktionelle Einordnung, kein Benchmark
As a rule of thumb:
- Latency and price are the bottleneck, the task is clearly scoped? Then Haiku.
- You want good quality at moderate volume? Then Sonnet.
- A long run where every intermediate step has to land? Then Opus.
Against Google’s Gemini Flash or GPT Luna, Haiku occupies the same volume zone: fast, cheap, moderate reasoning depth. Gemini Flash adds native video and audio, Luna trails Haiku slightly on coding. Which model fits is decided by the concrete modality of your task.
Read on
- Claude Haiku 4.5 — the current flagship with pricing and benchmarks
- LLM model families at a glance — Haiku next to GPT, Gemini and co.
- AI pricing explained — why the Haiku price matters for volume
- Building AI agents — Haiku as a parallel sub-agent
FAQ
- Its place in the line architecture: Haiku is tuned for low latency and low cost, Sonnet is the balanced mid-tier, Opus the top-tier line for the hardest reasoning and coding work.
- No. No model in the Haiku line supports effort levels; Anthropic's model spec explicitly lists "Default effort: Not supported" for Haiku 4.5. Reasoning depth runs instead through a manual, numeric thinking budget (budget_tokens).
- Tasks with high volume and low latency — classification, extraction, chat support — as well as a fast parallel sub-agent in multi-agent systems under an orchestrating Sonnet or Opus.
- As of September 2026, no. Haiku 4.5 from October 2025 remains the current flagship of the line, even though Sonnet, Opus and the new Fable/Mythos tier are already on generation 5.
- All three occupy the same volume zone: fast, cheap, moderate reasoning depth. Gemini Flash adds native video and audio, GPT Luna trails Haiku slightly on coding score.