Guides · 2026-10-13
Claude Haiku 5.5 pricing: the 100K-token cliff, costed
Anthropic released Claude Haiku 5.5 on 7 October 2026 at $0.10 per million input tokens and $0.50 per million output — a tenth of Haiku 4.5, and exactly the price of OpenAI’s GPT-6 Luna. The catch is one line on the pricing page: once a request’s prompt passes 100,000 tokens, the whole request is billed at five times those rates, cached tokens included. Below: the full rate card, what one request costs either side of the line, three monthly workloads priced against Luna, DeepSeek and Gemini Flash, and how to keep prompts under it.
The rate card
| Per million tokens | Haiku 5.5, prompt ≤100K | Haiku 5.5, prompt >100K | Haiku 4.5 (any size) |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache hit | $0.01 | $0.05 | $0.10 |
| Cache write, 5 min | $0.125 | $0.625 | $1.25 |
| Cache write, 1 hour | $0.20 | $1.00 | $2.00 |
| Batch input / output | $0.05 / $0.25 | $0.25 / $1.25 | $0.50 / $2.50 |
- The threshold counts everything you send. Anthropic defines prompt length as all input tokens of the request — new text, cache reads and cache writes. A 95K cached prefix plus a 6K question is a 101K prompt, and the cache hits on that request cost $0.05, not $0.01.
- It is per request, not per account. Each call is priced on its own; one long request does not reprice the rest of the month.
- 1M context, 128K output. The model accepts prompts up to a million tokens (API id
claude-haiku-5-5). The window is not the problem — the price above 100K is.
One request, either side of the line
Cost of 1,000 requests at each prompt size, with 1,000 output tokens each and no caching. Going from 100K to 101K tokens — 1% more prompt — takes Haiku 5.5 from $10.50 to $53.00, 5.0× the bill.
| Prompt | Claude Haiku 5.5 | GPT-6 Luna | DeepSeek V4.1 Flash (peak) | Gemini 3.8 Flash (2026 promo) | Claude Haiku 4.5 |
|---|---|---|---|---|---|
| 50K | $5.50 | $5.50 | $16.20 | $41.25 | $55.00 |
| 99K | $10.40 | $10.40 | $30.90 | $78.00 | $104 |
| 100K | $10.50 | $10.50 | $31.20 | $78.75 | $105 |
| 101K | $53.00 | $10.60 | $31.50 | $79.50 | $106 |
| 150K | $77.50 | $15.50 | $46.20 | $116 | $155 |
| 272K | $139 | $27.70 | $82.80 | $208 | $277 |
| 300K | $153 | $60.75 | $91.20 | $229 | $305 |
| 500K | $253 | $101 | $151 | $379 | $505 |
| 900K | $453 | $181 | $271 | $679 | $905 |
Tier rules: Claude Haiku 5.5 — 5× on the whole request above 100K; GPT-6 Luna — 2× input, 1.5× output above 272K; DeepSeek V4.1 Flash (peak) — flat; off-peak hours half price; Gemini 3.8 Flash (2026 promo) — flat; doubles on 1 Jan 2027; Claude Haiku 4.5 — flat. Cheapest per row in green.
Three things fall out of the table. Under 100K, Haiku 5.5 and Luna are the same price to the cent and the cheapest mainstream models in our data. Between 100K and 272K, Luna is five times cheaper, and DeepSeek V4.1 Flash — even at peak — undercuts Haiku 5.5 too. And above the line Haiku 5.5 is still half the price of Haiku 4.5 and below Gemini 3.8 Flash’s promo rate, so if you are upgrading from 4.5 there is no prompt size at which the new model costs more.
Three workloads, a month each
100,000requests a month per workload. Cache-write surcharges are left out for every model (they are small once a prefix is warm); Gemini’s cache storage fee is left out too.
| Workload | Claude Haiku 5.5 | GPT-6 Luna | DeepSeek V4.1 Flash (peak) | Gemini 3.8 Flash (2026 promo) | Claude Haiku 4.5 |
|---|---|---|---|---|---|
| Support chatbot 8K prompt (6K cached system prompt + tools), 400 tokens out | $46.00 | $46.00 | $112 | $345 | $460 |
| RAG Q&A 40K prompt (4K cached instructions + retrieved chunks), 800 out | $404 | $404 | $1,178 | $3,030 | $4,040 |
| Same RAG, retrieval too generous 110K prompt (4K cached), 800 out — only the chunk count changed | $5,520 | $1,104 | $3,278 | $8,280 | $11,040 |
| Long-document agent 150K prompt (120K cached document), 2K out | $2,600 | $520 | $1,212 | $3,900 | $5,200 |
The third row is the one to stare at. Nothing about the RAG app changed except how many chunks the retriever stuffed in: 40K to 110K tokens is 2.75× the input, but the Haiku 5.5 bill goes from $404 to $5,520 — 13.7×. On Luna the same change costs 2.7×. The long-document agent shows the other trap: its 120K cached document is cheap to re-read on Luna ($0.01 per million) but costs five times that on Haiku 5.5, because the cached tokens push the request over the line.
Staying under 100K
- Count before you send. Anthropic’s token-counting endpoint is free to call; check the prompt and trim (or reroute) when it is over 100,000. Trimming a 101K prompt to 99K takes 1,000 requests from $53.00 to $10.40.
- Cap retrieval by tokens, not by chunk count. A fixed top-k of 20 chunks is fine until someone indexes 8K-token PDFs. Budget the context (say 80K for chunks, the rest for instructions and history) and drop the lowest-ranked chunks first.
- Summarise long chat history instead of replaying it. Conversation length is the most common way a chatbot drifts over the line weeks after launch.
- Route long prompts elsewhere. If a share of requests genuinely needs 150K+ tokens, send just those to Luna (or DeepSeek off-peak) and keep Haiku 5.5 for the rest — the two are interchangeable on price below 100K.
- Batch what can wait. The Batch API halves both tiers; a long-prompt batch job at $0.25/$1.25 costs a quarter of Haiku 4.5’s standard $1/$5.
The app around the model
Even at Haiku 5.5 prices, the model is usually the bigger line on an AI feature’s bill. Our hosting data puts a small FastAPI backend at startup / small business traffic on these cheapest fitting plans:
- FastAPI on Orkestr: Starter, free tier
- FastAPI on DigitalOcean Droplets: Basic 512MB, about $4.00/month
- FastAPI on OVHcloud VPS: VPS-1, about $4.54/month
Against that, the RAG workload above costs $404 a month on Haiku 5.5 — and $5,520 once its prompts drift past 100K. Prompt length, not the server, is what to watch. A vector store and a response cache come next — see managed Postgres pricing (pgvector) and managed Redis pricing. If the front end is on Vercel, compare it with a container host first: Vercel vs Railway and Vercel alternatives by cost.
Should you switch?
- On Haiku 4.5 today: yes, on price alone — between 50% and 90% cheaper at the same token counts. Re-run your evals first; a new model is a new model.
- On GPT-6 Luna: price is a wash below 100K and a loss above it, so choose on quality. If you test Haiku 5.5, log prompt lengths before you route long requests to it.
- On Gemini 3.8 Flash: the promo price doubles on 1 January 2027 (3.7 and 3.5 Flash were already deprecated on 8 October). Haiku 5.5 is cheaper today at every prompt size and much cheaper next year.
- On DeepSeek: cheaper above 100K, more expensive below, and off-peak hours halve it again. The full field is in LLM API pricing in 2026.
Cost your own token volumes
The LLM API cost calculator uses the same data. For the hosting side, the hosting cost calculator ranks every provider for your stack and traffic. Running the chat route on Vercel? See what a streaming AI route costs in Fluid compute.
Prices verified 2026-10-13on Anthropic’s pricing and models pages, OpenAI’s pricing and GPT-6 Luna model pages, DeepSeek’s API pricing page and Google’s Gemini API pricing page. Base input/output prices come from our LLM data set (2026-10-12); hosting estimates from our hosting data set (2026-10-12). Token counts in the workloads are illustrative, not measured. No AI provider pays us anything, and none of the links in this article are affiliate links.
FAQ
How much does Claude Haiku 5.5 cost?
For prompts up to 100,000 tokens: $0.10 per million input tokens and $0.50 per million output tokens, with cache hits at $0.01 per million and cache writes at $0.125 (5-minute) or $0.20 (1-hour). For prompts over 100,000 tokens every rate is five times higher: $0.50 input, $2.50 output, $0.05 cache hits. The Batch API halves both tiers ($0.05/$0.25 and $0.25/$1.25).
Do cached tokens count toward Haiku 5.5’s 100K-token threshold?
Yes. Anthropic counts all input tokens of a request — fresh input, cache reads and cache writes — toward the prompt length. A request with a 95K cached prefix and 6K of new text is a 101K prompt and pays the higher rates on everything, including the cache hits. Each request is priced on its own, so earlier requests keep their price.
Is Claude Haiku 5.5 cheaper than GPT-6 Luna?
Up to 100K prompt tokens they cost the same: $0.10 input, $0.50 output and $0.01 cached input per million. Between 100K and 272K tokens Luna is five times cheaper, because Haiku 5.5 switches to $0.50/$2.50 while Luna keeps its short-context price. Above 272K Luna doubles its input and raises output by 1.5× ($0.20/$0.75), so Haiku 5.5 is still about 2.5× more expensive there.
Is Haiku 5.5 cheaper than Haiku 4.5 even above 100K tokens?
Yes. Haiku 4.5 is $1/$5 per million at any prompt size, so Haiku 5.5’s long-prompt rate of $0.50/$2.50 is still half of it. Below 100K tokens Haiku 5.5 is a tenth of the Haiku 4.5 price. There is no prompt size at which switching from 4.5 to 5.5 raises the bill, as long as the token counts stay the same.