tool notes · from issue #1
Claude Haiku 5.5: pricing, who it's for, and our take
· prices verified against Anthropic · covered in issue #1
Anthropic's cheapest, fastest small model — built for high-volume grunt work, not for running your whole agent solo. Here's the pricing, who should use it, and how we slot it into a real workflow.
What it is
Claude Haiku 5.5 (released 7 Oct 2026) is Anthropic's small-model tier: fast, cheap, and — for the first time in Haiku — with an adjustable effort setting so you can trade cost for a bit more intelligence. Anthropic positions it for high-volume, cost-sensitive work: summaries, compaction, classification, database lookups, live support, and especially as a subagent beside Sonnet 5.5 or Opus 5.5 on coding jobs.
API model id: claude-haiku-5-5. Also on AWS, Google Cloud, and Microsoft Azure / Foundry. See Anthropic's launch post and pricing docs.
Pricing verified 15 Oct 2026
Haiku 5.5 is priced by prompt length on the Claude API (USD per 1M tokens). A prompt over 100K tokens — counting all input including cache reads/writes — pays the higher tier:
| Tier | Input | Output | Cache read |
|---|---|---|---|
| Prompts ≤100K | $0.10 | $0.50 | $0.01 |
| Prompts >100K | $0.50 | $2.50 | $0.05 |
5-minute cache writes: $0.125 / $0.625. Batch API: 50% off input and output. Sources: Anthropic launch, platform pricing.
Compared with Haiku 4.5 ($1 / $5), that's roughly 90% cheaper on the ≤100K tier Anthropic says covers ~90% of prior Haiku traffic — and about 75% cheaper on average once token-use differences are counted.
Who it's for
Use it when…
- Multi-agent setups where subagents burn tokens on lookups, summaries, and "find every caller of X"
- High-volume, narrow tasks: classification, compaction, short CRM/doc Q&A
- Speed-sensitive loops (support, browser use) where latency matters more than depth
- You want adjustable effort on a small model without jumping to Sonnet prices
Skip it when…
- The job is complex agentic coding end-to-end — Anthropic itself says Sonnet 5.5 and Opus 5.5 still win there (e.g. Terminal-Bench-style work)
- You need one model to plan, refactor, and ship a big change alone
- Your prompts regularly blow past 100K and you haven't checked whether Sonnet + caching is cheaper for that shape
Our take
We covered Haiku 5.5 in Commit Weekly #1. The short version: yes, as a subagent — not as your only brain.
The workflow we actually recommend is plan big / grunt small / prove it:
- Plan with a frontier model — have it write PLAN.md (files, risks, definition of done) and wait for your OK.
- Hand the grunt work to Haiku 5.5 — summaries, call-site sweeps, extract-this-line jobs that would torch a Sonnet budget.
- Prove it at runtime — run real flows against the real stack; the test run says "done," not the agent.
Haiku 5.5 is the middle step. Cheap enough that you stop rationing subagents; capable enough that those subagents are useful. Keep Sonnet/Opus on the hard path.
Read the full issue write-up and the rest of that week's tools on issue #1, or browse the archive.
// subscribe
Get tools like this every week free
One email a week with the AI tools worth trying. Subscribing happens on Beehiiv, our newsletter platform.
Opens commits-newsletter.beehiiv.com. Unsubscribe anytime.
← Home · Issue #1 · Sources: anthropic.com/claude-haiku-5-5, platform pricing