Claude Haiku 5.5: 90% cheaper per token, but the new tokenizer moves the 100k price cliff
Anthropic released Claude Haiku 5.5 on October 7, 2026: 1M context, 128k max output, adaptive thinking with effort levels, and a list price of $0.10 per million input tokens, which is a tenth of what Haiku 4.5 costs. If you run classification, routing or extraction at volume, that headline is real money.
The headline isn't the whole bill, though. The newer tokenizer (the same one as Claude 4.7 and later) turns the same text into roughly 30% more tokens, and pricing steps up 5x at 100,000 prompt tokens. A prompt comfortably under 100k on Haiku 4.5 can land above it on 5.5. Here's the arithmetic, then the checklist.
What actually shipped
From Anthropic's announcement and the model docs: the model ID is claude-haiku-5-5 (Bedrock: anthropic.claude-haiku-5-5), with no date suffix, and retirement is no sooner than October 7, 2027. Context is 1M tokens (Haiku 4.5: 200k). Output goes up to 128k (Haiku 4.5: 64k), or 300k on Batches with a beta header. Adaptive thinking is on by default, controlled with output_config.effort; the overview page gives medium as the default.
The tokenizer produces "approximately 30% more tokens" for the same input text, with the exact figure depending on content. Images use the high-resolution tier, so a 2,000×1,500 image costs about 2.5x as many visual tokens. The minimum cacheable prompt drops from 4,096 to 512 tokens. There is no Priority Tier support, and no server-side fallback when a safety classifier returns stop_reason: "refusal".
The "What's new" page lists exactly five breaking changes. The migration guide goes further, with around a dozen items to handle.
Numbers that matter
| Per 1M tokens | Haiku 4.5 | Haiku 5.5 ≤100k prompt | Haiku 5.5 >100k prompt |
|---|---|---|---|
| Input | $1.00 | $0.10 | $0.50 |
| Output | $5.00 | $0.50 | $2.50 |
| Cache hit | $0.10 | $0.01 | $0.05 |
| Cache write (5m) | $1.25 | $0.125 | $0.625 |
| Input per 4.5-equivalent token (×1.3) | $1.00 | $0.13 | $0.65 |
The pricing page is explicit that the threshold applies to the whole request: prompt length "counts all of its input tokens, including cache reads and cache writes", and a request over 100k pays the higher rate on everything, cached part included.
Illustrative example 1: classifier, 1M requests/day. Each request has a 2,000-token prompt and a 50-token answer, counted in Haiku 4.5 tokens. On Haiku 4.5 that is 2.0B × $1 + 50M × $5 = $2,000 + $250 = $2,250/day. On Haiku 5.5 (+30%) it is 2.6B × $0.10 + 65M × $0.50 = $260 + $32.50 = $292.50/day, a cut of about 87%. Adding 300 thinking tokens per request (my assumption, since thinking is on by default) adds 300M × $0.50 = $150, for $442.50/day. That is still about 80% cheaper.
Illustrative example 2: the cliff. Here the prompt is 80,000 tokens with a 1,000-token answer, again in Haiku 4.5 tokens. On Haiku 4.5: 80k × $1/M + 1k × $5/M = $0.085. On Haiku 5.5, 80k × 1.3 = 104k, which is over the threshold, so 104k × $0.50/M + 1.3k × $2.50/M = $0.052 + $0.00325 = $0.055. That is 35% cheaper, not 90%. The same job at 76k old tokens becomes 98.8k new tokens, under the threshold: $0.00988 + $0.00065 = $0.0105.
The break-even sits at about 100,000 / 1.3 ≈ 77k Haiku-4.5 tokens. Above it, a ~5% longer prompt costs ~5x more. The output figures assume generated text inflates by the same ~30% as input. Anthropic only documents the input figure, so treat that part as my estimate.
Above about 77k Haiku-4.5 tokens, a ~5% longer prompt costs ~5x more.
Anthropic's own footnote is the clearest summary. List prices fall 90% (≤100k) and 50% (>100k), yet the average claimed saving is "around 75%" once the request mix and the tokenizer are included. About 90% of Haiku 4.5 requests were under 100k. The 15-point gap between the list cut and the claimed average is mostly the tokenizer and the cliff.
The trade-offs that bite
Agent loops cross 100k faster than you think. Haiku 4.5 kept only the latest turn's thinking blocks. On Haiku 5.5, thinking blocks from all earlier assistant turns stay in context and count as input. A subagent that stayed under 100k on 4.5 can drift over it mid-loop on 5.5, and every later turn pays the long rate.
Effort is a cost dial, not a quality knob. Artificial Analysis measured about 162k output tokens per Intelligence Index task at max effort, roughly 3x GPT-6 Luna at max. At high effort it measured about 55k tokens per task and a score of 38, against 43 at max. Pick the effort level from your eval, not from the default.
The API is stricter. These requests now return a 400: budget_tokens thinking, assistant prefill, and computer_20250124. Sampling parameters should be removed. Thinking blocks are bound to the account that produced them, and the API silently drops blocks from another account. Changing earlier turns before a replayed thinking block can also fail. The same mechanics came with Sonnet 5.5 and Opus 5.5; see the Sonnet 5.5 migration notes and the Opus 5.5 checklist.
What I'd do
I'd re-tokenize a production sample first: run real prompts through token counting on claude-haiku-5-5 and plot the distribution against 100k. Anything in the 70k to 100k band on Haiku 4.5 needs a decision. Then I'd cap prompts deliberately. If you are near the line, trim retrieval to stay under ~95k new tokens, or accept the long tier knowingly; context editing and thinking-block clearing are the tools for agent loops.
Recompute max_tokens, because thinking counts toward it, and a tight cap tuned for 4.5 can end in stop_reason: "max_tokens" before any text appears. Replace prefill with structured outputs or enum-constrained tools, and select content blocks by type, never by index. Handle refusal explicitly: with no server-side fallback, your router has to catch it. Finally, set effort per route and re-check image pipelines. Use low for classification and medium or high for extraction. Visual tokens rise ~2.5x for typical photos.
What I couldn't verify
The lead's "1M context at $0.10" framing is misleading. $0.10 applies only to prompts ≤100k; above that, input is $0.50. The ~30% figure is Anthropic's average, and I have not measured it on my own corpora; code, non-English text and JSON may differ. Benchmarks on the announcement are self-reported. Terminal-Bench 4.0 is listed at 39.2% by Anthropic and 33% by Artificial Analysis, and AA also says its cost figures don't yet include the tiered pricing. I found no published worked calculation behind Anthropic's "75%" claim.
I'll revisit with token-count deltas from real client traffic.
Sources
Primary: Anthropic, Claude Haiku 5.5; Claude Docs, Claude Haiku 5.5 overview; Claude Docs, What's new in Claude Haiku 5.5; Claude Docs, Migrating to Claude Haiku 5.5; Claude Docs, Pricing. Corroboration: Artificial Analysis, Claude Haiku 5.5.