Claude Haiku 5.5 costs 5x past 100K tokens: the math

Model launch · 2026-10-10 Claude Haiku 5.5 costs 5x past 100K tokens claude-haiku-5-5Stable $0.10 / $0.50 Lazu

Short answer: Claude Haiku 5.5 costs $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens. Over 100,000 tokens, every price for that request (input, output, cache reads and writes) is 5x. One token over the line can make the whole request cost five times as much.

claude-haiku-5-5Anthropic · 1M · Stable $0.10 / $0.50Model page →

USD per million tokens

LaneInputOutputCache readCache write
Anthropic official price$0.10$0.50$0.01$0.125 (5m)
Input over 100k$0.50$2.50$0.05$0.625 (5m)
Lazu Stable · Same as official$0.10$0.50$0.01$0.125 (5m)
Input over 100k$0.50$2.50$0.05$0.625 (5m)

A request whose input (cached tokens included) passes the threshold is billed entirely at the long-context prices.

A dash means no price is recorded, not free usage. The request receipt is the final billing record.

The rule #

From Anthropic's pricing documentation:

  1. The tier depends on prompt length, and prompt length counts all input tokens, including cache reads and cache writes. A cache hit still counts.
  2. The whole request changes price, not only the tokens past the line. Input, output, 5-minute cache writes, 1-hour cache writes and cache reads all switch to 5x together.
  3. Exactly 100,000 tokens is still the lower tier. Only prompts over it pay more.
  4. Output tokens don't count toward the threshold, but in the higher tier they are billed at the higher rate.
  5. Each request is priced on its own. Earlier requests keep their prices.
Item Prompt ≤ 100K Prompt > 100K
Input $0.10 $0.50
Output $0.50 $2.50
Cache read $0.01 $0.05
Cache write (5 minutes) $0.125 $0.625
Cache write (1 hour) $0.20 $1.00

USD per million tokens. Lazu's stable lane matches the official price.

Worked examples #

Case Math Cost
50K prompt, 2K output 50K × 0.10 + 2K × 0.50 ≈ $0.006
Exactly 100,000 prompt, 2K output 100K × 0.10 + 2K × 0.50 ≈ $0.011
100,001 prompt, 2K output 100,001 × 0.50 + 2K × 2.50 ≈ $0.055 (5x)
150K prompt (140K cache hit), 2K output 140K × 0.05 + 10K × 0.50 + 2K × 2.50 ≈ $0.017
Same task kept under 100K (90K cached + 10K new), 2K output 90K × 0.01 + 10K × 0.10 + 2K × 0.50 ≈ $0.003

The last two rows are the same kind of task: keeping the context under 100K costs about a sixth of crossing the line.

Still cheaper than Haiku 4.5 #

Haiku 4.5 is $1 / $5 with no tiers. Even past 100K, Haiku 5.5's prices ($0.50 / $2.50) are half of Haiku 4.5's.

The tokenizer matters too. Anthropic says Haiku 5.5 uses a newer tokenizer that produces about 30% more tokens for the same text (it depends on the content). Redo the 150K example with that ratio and it is about $0.022 on Haiku 5.5 against about $0.034 on Haiku 4.5, so Haiku 5.5 still wins.

The comparison that matters is not 5.5 against 4.5; it is 5.5 under 100K against 5.5 over it.

Why long sessions cross the line #

  • The context window is 1M. In Claude Code and other agents, a long session keeps growing past 100K.
  • Earlier thinking stays in context. Anthropic notes that Haiku 5.5 keeps thinking blocks from all earlier turns and counts them as input; Haiku 4.5 kept only the latest. Multi-turn input grows faster as a result.
  • Cache hits still count toward length. Caching makes tokens cheaper; it does not make the prompt shorter.

Staying under 100K #

  1. Give Haiku 5.5 short jobs. Subagents, background tasks and single-file edits suit it better than being the main model of a long session. In Claude Code, put it in ANTHROPIC_DEFAULT_HAIKU_MODEL and keep Sonnet 5.5 as the main model.
  2. Split the work. Send only the context this step needs, not the whole history.
  3. Use auto-compaction. Claude Code's CLAUDE_CODE_AUTO_COMPACT_WINDOW can go as low as 100000. It limits how far a session grows but does not keep every request under 100K.
  4. Cache small pieces. Haiku 5.5's minimum cacheable prompt is 512 tokens (4,096 on Haiku 4.5), so fixed system prompts and tool definitions cache more easily.
  5. Watch each request's usage. input_tokens + cache_creation_input_tokens + cache_read_input_tokens in the response's usage is roughly the prompt length the tier uses. That is our reading; Anthropic's documentation is authoritative.

Check it on Lazu #

curl --fail-with-body https://api.lazu.ai/v1/messages \
  -H "Authorization: Bearer $LAZU_API_KEY" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{"model":"claude-haiku-5-5","max_tokens":64,"messages":[{"role":"user","content":"Reply with only OK."}]}'

The X-Lazu-Request-Id response header identifies the request. Open it with GET /api/usage/requests/{request_id} or in the console's usage log to see its input, cache and output tokens and the actual charge.

Next #

Related
2026-11-18OpenCode custom provider: Claude, GPT and DeepSeek on Lazu 2026-11-18Claude Code gateway contract: 10 traps that stall agents 2026-10-28BAAI/bge-m3 embeddings API: 100+ languages, $0.02 per 1M

Try it with a key

Start free