Playbooks · Coding tasks
Haiku 5.5 is 90% cheaper only under 100K tokens: price the tier jump and the tokenizer before you route background work to it
Written November 1, 2026 · Edited November 4, 2026
For a builder routing summaries, compaction, classification, or subagent work to the cheapest Claude model: the headline price stops at 100,000 prompt tokens, the higher price then applies to the whole request, and the same text counts about 30% more tokens than it did on Haiku 4.5. Work out the price of one real task, in Haiku 5.5 tokens, before moving a queue.
What changed
Anthropic released Claude Haiku 5.5 on 7 October 2026, describing it as built for high-volume, cost-sensitive work such as summaries, compactions, database queries, and classification, and as a subagent alongside the larger models in the Claude 5.5 family. It has a 1-million-token context window and up to 128,000 output tokens, up from 200,000 and 64,000 on Haiku 4.5, according to Anthropic's what's-new page. It is the first Haiku model with an adjustable effort setting, which trades response quality against speed and cost.
Anthropic's announcement says Haiku 5.5 costs around 75% less to run on average than Haiku 4.5. That average is not the sticker price. The sticker price, 90% below Haiku 4.5, holds only for requests whose prompt is up to 100,000 tokens. Anthropic also says prompts up to 100,000 tokens made up around 90% of requests to the previous Haiku model, which is the main reason the average sits where it does. Averages across other people's traffic do not price your tasks. The rest of this page prices the two effects that move your bill away from the headline: the tier jump past 100,000 tokens, and the new tokenizer.
The two price lists
Anthropic's pricing page lists two rows for one model. For prompts up to 100,000 tokens: input $0.10, output $0.50, five-minute cache write $0.125, one-hour cache write $0.20, and cache read $0.01, all per million tokens. For prompts over 100,000 tokens: input $0.50, output $2.50, five-minute cache write $0.625, one-hour cache write $1.00, and cache read $0.05, all per million tokens. Every line is five times the lower-tier figure. The Batch API halves input and output in both tiers.
The jump is a step, not a slope. A request whose prompt is over 100,000 tokens pays the higher prices on the whole request, including its output and its cache reads, not only on the tokens past the line. Each request is priced on its own: a request over the threshold pays the higher prices even when part of its prompt is a cache hit, and earlier requests keep the prices they were billed at. Prompt length counts all input tokens, including cache reads and cache writes, so a prompt that looks safely under the line in fresh tokens can cross it once cached tokens are counted. Haiku 4.5 billed $1.00 input and $5.00 output per million tokens, so the lower tier is 90% below it and the upper tier is 50% below it, before the tokenizer effect below.
The tokenizer hands part of the discount back
Haiku 5.5 uses the same newer tokenizer as the later Claude models, and Anthropic's model overview states that the same text counts as approximately 30% more tokens than on Haiku 4.5, with the exact increase depending on the content. The what's-new page repeats the figure and tells developers to recount prompts and revisit maximum-output limits and cost estimates. A prompt that measured about 77,000 tokens on Haiku 4.5 measures close to 100,000 on Haiku 5.5. Images move further: an image of 2,000 by 1,500 pixels costs about 2.5 times as many visual tokens on Haiku 5.5 as on Haiku 4.5, according to the migration guide, because Haiku 5.5 uses the high-resolution image tier.
Multi-turn conversations carry a second token increase that the tokenizer figure alone does not explain. Haiku 5.5 keeps thinking blocks from all earlier assistant turns in context, and they count as input tokens, while Haiku 4.5 kept only the latest turn's blocks. A long agent loop therefore grows its prompt faster than a single-turn recount would suggest. The model identifier is claude-haiku-5-5.
Worked examples: price one real task
All three examples use the published rates above and the approximate 30% recount. They are arithmetic, not measurements of any particular workload. Count your own prompts with the model set to Haiku 5.5, as the migration guide instructs, rather than reusing Haiku 4.5 counts.
Example 1: a short classification task, under the line. Suppose a task measured 10,000 input tokens and 500 output tokens on Haiku 4.5. On Haiku 4.5 it costs 10,000 × $1.00 plus 500 × $5.00 per million, which is $0.0125. Recounted at 30% more, the input is about 13,000 tokens. On Haiku 5.5, in the lower tier, it costs 13,000 × $0.10 plus 500 × $0.50 per million, which is $0.00155. That is 87.6% less than Haiku 4.5, not 90%, because the tokenizer added tokens before the cheaper rate was applied.
Example 2: a long agent prompt, over the line. Suppose a prompt measured 90,000 input tokens on Haiku 4.5, with 1,000 output tokens. Recounted, it is about 117,000 tokens on Haiku 5.5, past the threshold. At the upper-tier rates it costs 117,000 × $0.50 plus 1,000 × $2.50 per million, which is $0.061. The same 117,000-token prompt at the lower-tier rates would cost $0.0122. The extra 17,000 tokens past the line do not cost a little more; they move the whole request to rates five times higher. On Haiku 4.5 the same task cost $0.095, so the saving in this case is about 36%, not 90% and not 75%.
Example 3: a prompt that only crosses the line with its cache. Suppose a request carries 80,000 fresh input tokens and 30,000 tokens read from the cache. Its prompt length is 110,000 tokens, so it pays the upper tier on everything, including the cache reads at $0.05 per million instead of $0.01. A budget built on fresh tokens alone would have placed it in the lower tier. Count cache reads and cache writes in the prompt length, as the pricing page states, and test the boundary with a real request rather than assuming a prompt sits under it.
Effort changes the bill, with no published multiplier
The effort setting in Haiku 5.5 has five levels: low, medium, high, xhigh, and max, with medium as the default. Anthropic's prompting guidance says low is cheapest and fastest, suited to short and high-volume work, while higher levels suit longer agent tasks and strict instruction following, and that thinking and replies get much longer at xhigh and max. Anthropic publishes no cost multiplier per effort level. Do not budget effort as a fixed percentage. Run the same task at two or three effort levels on your own evaluation, count the tokens each level actually uses, and compare the result against Sonnet 5.5 on the same task, as the prompting guidance suggests for the highest levels. Anthropic's own positioning, in the launch announcement, is that the larger models remain the better choice for complex agentic coding tasks, and Haiku 5.5 is best suited to more narrowly scoped tasks such as compaction, summarization, or subagent work.
What to do about it
- Recount before you route. Take a representative sample of real prompts, count them with the model set to Haiku 5.5, and include cache reads, cache writes, and thinking blocks from earlier turns. A Haiku 4.5 token count is not a Haiku 5.5 token count.
- Split the queue at 100,000 tokens. Price tasks under the line and tasks over the line separately, using the two rows in the pricing table. A single blended rate hides the cliff. If a task sits near the line, test whether trimming context, summarizing earlier turns, or splitting the work keeps it under the line, and price the trimming work too.
- Set a per-task ceiling, not a per-token hope. Decide the maximum acceptable cost of one task at the upper-tier rates, and route to Haiku 5.5 only the tasks that stay under it at their recounted size. Tasks that regularly cross the line need a separate decision: a larger model, a shorter prompt, or a different split of the work.
- Revisit output limits. Thinking tokens count toward the maximum output limit, so a limit tuned for Haiku 4.5 can stop after a thinking block and before any text. Revisit the limit alongside the cost estimate, as the what's-new page advises.
What can go wrong
- Budgeting the headline rate for a long-prompt queue. Agent loops, compaction, and subagent work with large contexts are exactly the workloads that cross 100,000 tokens. The headline rate describes the other 90% of the previous model's traffic, not every queue.
- Reusing Haiku 4.5 token counts. The same text is about 30% more tokens, images cost more, and earlier thinking blocks now stay in context. Each of those pushes a prompt toward the line.
- Counting only fresh tokens. Prompt length includes cache reads and cache writes. A prompt that is under the line before caching can be over it after.
- Assuming a higher effort level costs a fixed amount more. No multiplier is published. The only sound budget for effort is a measured one, on your own tasks.
Limits of this evaluation
- Every price here is a list price from Anthropic's pricing page, in US dollars, for the Claude Platform. Cloud platforms that bill through a provider invoice separately, and any negotiated discount, priority tier, or data-residency multiplier changes the bill. Priority Tier is not supported on Haiku 5.5, according to the migration guide.
- The 75% average saving is Anthropic's figure across its own traffic, built on the tier split and the tokenizer change. It is not a promise for any particular queue, and the worked examples above show two queues where the saving is 88% and 36%.
- No capability claim is made here beyond Anthropic's own description of the tasks Haiku 5.5 is built for. Benchmark figures from the launch announcement are not repeated, because a benchmark score does not price a task.
Related playbooks
- Check whether your Claude Code hooks are adding tokens to every prompt or cutting them
- Haiku 5.5 is now the default Haiku
- Measure what Claude Code's background-agent summaries cost you, and decide what to turn off
- 180M repos: most agent work is unattributed
Sources
- https://www.anthropic.com/claude-haiku-5-5
- https://platform.claude.com/docs/en/about-claude/pricing
- https://platform.claude.com/docs/en/models/haiku-5-5/overview
- https://platform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5
- https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide
Back to: Coding playbooks hub · Also: Token cost calculator · What an agent costs · Pick the cheapest model · Measure background-agent summary cost