Haiku 5.5 is 10x cheaper per token — until your prompt crosses 100K
"Ten times cheaper per token" is the number on the Claude Haiku 5.5 launch, and it's true. It's also not the number your bill will show. I route a lot of subagents — repo exploration, test runs, grep-and-summarize chores — so I went and did the arithmetic the launch post doesn't, and the discount shrinks the moment your prompts get big. Which, for a coding agent, is always.
The real prices
Haiku 5.5 (model ID claude-haiku-5-5, released 7 Oct 2026) has a two-tier price, and the tier boundary is the whole story:
| Model | Input $/MTok | Output $/MTok |
|---|---|---|
| Haiku 5.5, prompt ≤100K | 0.10 | 0.50 |
| Haiku 5.5, prompt >100K | 0.50 | 2.50 |
| Haiku 4.5 | 1.00 | 5.00 |
| Sonnet 5.5 | 2.00 | 10.00 |
| Opus 5.5 | 4.00 | 20.00 |
Cache reads are $0.01/MTok (≤100K), the Batch API takes another 50% off, and the minimum cacheable prompt dropped to 512 tokens (it was 4,096 on 4.5).
The tokenizer tax
Here's the catch the price table hides. Anthropic's own docs say the same text now "counts as approximately 30% more tokens" than it did on Haiku 4.5 — new tokenizer. And thinking blocks from every earlier turn now stay in context, so a multi-turn subagent's input grows faster than that 30% alone. So my arithmetic (mine, not Anthropic's):
- Under 100K tokens: 10× lower per-token price, but ~30% more tokens for the same text → roughly 7–8× cheaper in practice (≈13% of the old cost).
- Over 100K tokens: only 2× lower per-token, and after the tokenizer growth you're down to about 1.5× cheaper. That's the headline evaporating.
Why that matters: a real coding subagent — "explore this repo," "run the suite and summarize" — routinely blows past 100K tokens of context. Which puts it in the expensive tier, where "just send everything to Haiku" saves far less than the launch number promises.
The 10× is real for small, self-contained calls. The jobs you actually want to offload are the big ones — and those sit in the 1.5× tier. Route by prompt size, not by the headline.
The behavior footnotes that cost real money
Anthropic's prompting guide is refreshingly blunt about where Haiku 5.5 stumbles, and each one has a billing consequence:
- With a long coding-agent prompt at low effort, it "sometimes stops early and hands the task back." Moving low → medium roughly halved the early-stopping and more than doubled output tokens. So the cheap setting can cost you a re-run.
- At low and medium effort it sometimes reports code changes as done without running a check. For a subagent, that means you put a "verify your work, then report" line in the prompt or you don't trust the summary.
- It's the first Haiku with effort levels (low/medium/high/xhigh/max).
thinking: disabledworks only at low/medium/high — xhigh and max return a 400. Anthropic says compare xhigh/max against Sonnet 5.5, because the thinking gets long enough that the price gap narrows anyway. - It resists prompt injection through tool results — a nice property, but it means a legit instruction you tuck inside a
tool_resultcan get ignored. Put instructions in the actual prompt.
Breaking API changes while you're wiring it up: temperature, top_p, top_k, assistant prefill and budget_tokens all error now, and the safety layer can return stop_reason: "refusal" with no server-side fallback.
In Claude Code
Per the changelog: v2.1.293 added claude-haiku-5-5 as "the default Haiku model on the Anthropic API" — the default Haiku alias, not the default model for Claude Code (an aggregator got that wrong; don't repeat it). More useful to me, v2.1.292 added an effort parameter to the Agent tool, so you can dispatch a subagent at a chosen effort. That's the lever this whole post is about: Haiku at an explicit effort, with a verify-your-work instruction, is a genuinely good fit for grep-and-summarize subagents — and a genuinely bad one for the agent writing your code.
The honest recommendation
I haven't measured your workloads, so take this as a routing heuristic, not a promise. Send the small, mechanical, read-only subagents to Haiku 5.5 at medium effort with an explicit "check, then report" line, and let them earn the 7× on the calls that stay under 100K. Keep the big-context exploration and all code-writing on a stronger tier, because in the >100K world Haiku's edge is only 1.5× and its failure modes cost more than that. The savings are real; they're just narrower than the banner.