Claude Haiku 5.5
A million-token context at $0.10 per million input tokens — long-context work stops being a budget decision.
Anthropic’s small, fast model. “Small” is about price and speed rather than a ceiling you would hit on everyday work. The headline is the combination: a million tokens of context — roughly a 700-page book — at ten cents per million input tokens.
What you build with it
- Put a whole codebase in one prompt instead of building retrieval for it
- Run the high-volume, low-difficulty steps of a pipeline: routing, classification, extraction, validation
- Summarise long documents you would otherwise have to chunk
- Back a chat feature where latency matters more than deep reasoning
What it gets you
- Long context stops being a budget decision
- Fast enough to sit in an interactive loop
- Stacks with prompt caching for another large discount
- Cheap enough to run speculatively and throw the result away
- Context
- 1,000,000 tokens
- Input
- $0.10 / 1M tokens
- Output
- $0.50 / 1M tokens
- Class
- Small / fast tier
Straight from the source
Why it matters
The interesting number is not the context window, it is the context window multiplied by the price. A million tokens at ten cents means you can stop building clever retrieval for mid-sized corpora and just put the corpus in the prompt. For a lot of internal tools, that is a simpler system that works better.
When a big cheap context beats a RAG pipeline
A RAG pipeline is several moving parts: a chunker, an embedding model, a vector store, a retriever, and a reranker — each with its own failure mode, and all of them upstream of the answer. When retrieval picks the wrong chunk, the model cannot recover, and you debug it by reading embeddings.
If the whole corpus fits in a million tokens, you can delete all of it. A full codebase, a quarter of support tickets, a long specification plus its revision history — these now fit, and at this price a full-context call costs about ten cents.
This stops working somewhere above a million tokens, and attention still degrades across very long inputs. But the threshold where "just send everything" wins has moved a long way up.
Where the cheap tier earns its place in a pipeline
In a multi-agent system most calls are not the hard call. Routing, classification, extraction, validation, summarising one step for the next — these are the majority of the traffic and almost none of the difficulty.
Running those on a small model and reserving the frontier model for the genuinely hard step is the single largest cost lever in an agent pipeline, and it usually costs nothing in quality.
That is the same hybrid-LLM pattern as routing between a local Ollama model and a hosted API — the cheap model handles the volume, the expensive one handles the judgement.
Prompt caching compounds it
If a long prefix is stable across calls — a system prompt, a schema, a document everyone asks about — caching it means you pay full price once and a fraction afterwards.
A long-context model and prompt caching are the same optimisation seen twice: put the stable bulk at the front, keep the variable part small, and only the variable part costs full price on each call.
Reach for it when
- A corpus fits in a million tokens and you would rather not run a retrieval stack
- High-volume, low-difficulty calls dominate your pipeline
- A long stable prefix is reused across many requests (pair it with prompt caching)
Watch out for
- A small model is still a small model — keep the hard reasoning step on a frontier tier
- Attention quality degrades over very long inputs; a million tokens of context is not a million tokens of attention
- Pricing and limits move — check the current model card before you budget on it