The fastest, cheapest Claude — now 90% cheaper than Haiku 4.5, with a 1M-token context window for shorter prompts.
Claude Haiku 5.5 bills $0.10 / 1M input and $0.50 / 1M output for prompts up to 100K tokens. Beyond 100K tokens, the rate rises to $0.50 / 1M input and $2.50 / 1M output.
Anthropic released Claude Haiku 5.5 on October 7, 2026, calling it a significant step up over Haiku 4.5 — and the pricing backs that up: $0.10 input / $0.50 output per 1M tokens (for prompts up to 100K) is roughly 90% cheaper than Haiku 4.5's $1/$5 rate. It also grows the context window from 200K to 1M tokens, the largest jump any Haiku generation has made. Anthropic kept Haiku 4.5 listed too, rather than retiring it, since the two now ship at very different prices.
The one wrinkle: Haiku 5.5 uses the newer tokenizer introduced with Claude 4.7, which tends to produce somewhat more tokens than Haiku 4.5's tokenizer for the same input text — so raw per-token prices understate the real-world discount slightly less than the headline numbers suggest, though it's still a steep cut either way.
Haiku 5.5 is built for the same high-volume, latency-sensitive work Haiku always targeted — classification, extraction, routing, templated generation — but now with enough context to handle much larger documents or longer conversation histories before stepping up to a bigger model. Prompts beyond 100K tokens bill at a higher rate ($0.50/$2.50), which keeps the economics sane for the common case while still supporting long-context jobs when needed.
As with prior Haiku generations, step up to Sonnet 5.5 or beyond for tasks that need real deliberation — legal nuance, subtle code review, multi-step research synthesis. Haiku optimizes for fast and cheap, not for depth.
| Model | Input / 1M | Output / 1M | Context |
|---|---|---|---|
| Claude Haiku 5.5 (≤100K) | $0.10 | $0.50 | 1M |
| Claude Haiku 4.5 (legacy) | $1 | $5 | 200K |
| GPT-6 Luna | $0.10 | $0.50 | 1M |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 1M |
| DeepSeek V4.1-Flash (off-peak) | $0.15 | $0.60 | 1M |
Haiku 5.5's price cut puts it right in line with budget models from other labs — it now ties GPT-6 Luna on input and undercuts Gemini Flash-Lite and DeepSeek's Flash tier on output, while still keeping Claude's instruction-following and drop-in compatibility with a Sonnet/Opus/Fable stack. For a Claude-based pipeline, there's little reason left to stay on the pricier Haiku 4.5 unless you specifically need its older tokenizer's token-counting behavior.