My subagents are already set to haiku. Will my bill automatically drop 90% after the upgrade?
Not automatically. "90% cheaper" applies only to prompts up to 100K tokens, and it is the nominal unit price. The new tokenizer counts about 30% more tokens, Adaptive Thinking is on by default and adds output, and Anthropic's own estimate of the average reduction is about 75%. If your Subagent prompts often exceed 100K, the rate is five times higher, so the real bill depends on your prompt-length distribution, not just the unit price.
Should I switch every Subagent to Haiku 5.5?
It fits summarization, classification, retrieval, Compaction and high-volume document Q&A. For complex agentic coding, Anthropic says Sonnet 5.5 or Opus 5.5 remain the better choices. A safer approach is to take a batch of real tasks, run them on Haiku 5.5, spot-check the output, and escalate to a stronger model only when a check fails, rather than switching everything at once. beam.ai's worked example even includes a second Haiku verification pass in the cost, and it adds only about $0.031.
How do I avoid crossing the 100K threshold by accident?
A few direct measures: retrieve only the relevant passages instead of handing a whole document to the Subagent; keep the Orchestrator's brief short; ask the subagent to return only the fields you need in a structured format; for long multi-turn subagents, condense the context periodically, since Haiku 5.5 keeps earlier thinking blocks by default; and cache the orchestrator's stable System Prompt and tool definitions. Whether cached reads count toward the 100K is not clearly stated, so assuming they do is the safe choice.
I'm not an engineer; I only track the team's AI tool budget. What should I watch in this news?
Don't use the headline "90% cheaper" to estimate next month's bill. Ask the team for a week of real usage and see how many requests have prompts above 100K, because that portion is billed at five times the rate. Also, after the model change, Token counts shift because of the tokenizer and the default thinking behavior, so treat the first month's bill as a new baseline rather than comparing it directly with the old one.
On October 7, 2026, Anthropic released Claude Haiku 5.5 (model ID claude-haiku-5-5), and Claude Code v2.1.293, shipped the same day, lists it as the default Haiku model. Anthropic calls it its "cheapest and fastest small model to date," aimed at summaries, Compaction, classification and subagent work. For anyone who uses Haiku as a Subagent or background model, this is a generation change in both price and behavior, not just a new label.
Published specs: a 1M-Token context window, 128K max output (up to 300K for batch jobs in beta), text and image input, and a June 2026 Knowledge Cutoff. Pricing per million tokens comes in two tiers:
Against Haiku 4.5 at $1 / $5, prompts up to 100K are 90% cheaper. But the new tokenizer counts about 30% more tokens for the same text, and Anthropic's own estimate, after adjusting for that, is roughly 75% cheaper on average, not 90%.
According to beam.ai's write-up, once a request crosses 100K the higher rate applies to the whole request, not only to the tokens above the line (check the wording on the official pricing page yourself). Its worked example: an 8,000-token subagent call costs about $0.00155, while the same kind of task at 120,000 tokens costs about $0.064, roughly forty times more. That is the author's estimate, not an official figure, but the direction is clear.
One more detail that is easy to miss: Haiku 5.5 keeps previous thinking blocks in context by default, so a subagent with many back-and-forth turns can quietly cross 100K without ever reading a large document. Whether cached-read tokens count toward the 100K is not clearly stated; beam.ai advises assuming they do.
Anthropic positions Haiku 5.5 as a subagent under Opus 5.5 or Sonnet 5.5, not as the lead model for complex agentic coding, and says the latter two remain the better choices for complex coding tasks. beam.ai's illustrative run (the author's own assumptions) shows an Opus 5.5 Orchestrator with Haiku 5.5 subagents costing about $0.445 for a full run, with the subagents only about 7% of it ($0.031); the same run entirely on Opus 5.5 costs about $1.65. The savings sit in the subagent portion, but the hand-off cost, with the orchestrator writing a brief and reading the answer, is about 4.8 times the subagent's own work, so the shorter the brief and the more structured the reply, the more you keep.
Separately, v2.1.292 added an effort parameter to the Agent Tool, which controls how hard a subagent works. Haiku 5.5 is also the first Haiku-class model with adjustable effort, defaulting to medium.
Temperature, top_p or top_k returns a 400. If a Haiku 4.5 caller sets any of them, remove them first.The benchmark numbers are Anthropic's own: Terminal-Bench 4.0 at 39.2% (Haiku 4.5: 0.0%) and OSWorld 2.1 offline subset at 72.4% (Haiku 4.5: 15.7%). Independent Artificial Analysis scores its Intelligence Index at 43 on maximum effort and 34 on medium, against 17 for Haiku 4.5 with reasoning. Maximum effort produced far more output tokens than medium (the write-up cites 440 million versus 54 million), so for high-volume subagent work medium or high is usually the better trade. Also, its safety classifiers can decline requests, and there is no server-side fallback model.