Claude Haiku 5.5: The Cheap Layer for Triage, CRM Lookups, and Subagents

Claude Haiku 5.5: The Cheap Layer for Triage, CRM Lookups, and Subagents

On October 7, 2026, Anthropic released Claude Haiku 5.5, calling it "the cheapest, fastest, and most capable small model we've ever released." The target is not another general chatbot. It is high-volume, cost-sensitive work: summaries, compactions, database queries, classification, live customer support, and browser use.

If you already run agents that call Claude dozens or hundreds of times a day, that is the layer that hits your bill first. The question is whether Haiku 5.5 is good enough for triage and lookups so you can keep Sonnet and Opus for the hard steps.

What Anthropic shipped

Anthropic says Haiku 5.5 handles quick, repetitive workloads and pairs with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. It is also positioned as Anthropic's fastest model to date at each model's standard speed, with a footnote that Opus Fast Mode can still be quicker.

Developers call it on the Claude Platform as claude-haiku-5-5. Anthropic says it is available on Amazon Web Services, Google Cloud, and Microsoft Azure. Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting, so you can trade cost against intelligence on the same model id.

The price cut that matters

Anthropic states that, on average, Haiku 5.5 costs around 75% less to run than Haiku 4.5. The footnote behind that average is more useful for budgeting:

• Priced 90% lower than Haiku 4.5 for requests up to 100,000 tokens
• Priced 50% lower for requests over 100,000 tokens
• On Haiku 4.5, 90% of requests fell into the up-to-100k category

Anthropic's announcement table lists Haiku 5.5 prices per 1 million tokens for prompts up to / over 100k as:

• Input: $0.10 / $0.50
• Output: $0.50 / $2.50
• Cache reads: $0.01 / $0.05
• Cache writes: $0.125 / $0.625

The same table lists Haiku 4.5 at $1.00 input and $5.00 output per million tokens.

That is list pricing from Anthropic, not a promise that your agent stack will drop 75%. Tokenizers, retries, and effort settings change real spend. Anthropic also notes that Haiku 5.5 uses a newer tokenizer and can use slightly more tokens per task than Haiku 4.5.

Alongside the launch, Anthropic cut Sonnet 5.5 cache reads to $0.10 per million tokens from $0.20, and says that reduces Sonnet 5.5 cost on most agentic tasks by around 20%. Max and Team subscribers also get new monthly API credits: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team. Those credits are Anthropic's stated rollout, not a third-party audit of your invoice.

Benchmarks, framed as vendor numbers

Anthropic published large jumps on computer use and agentic coding versus Haiku 4.5:

• OSWorld 2.1 Offline subset: Haiku 5.5 at 72.4%, Haiku 4.5 at 15.7%
• Terminal-Bench 4.0: Haiku 5.5 at 39.2%, Haiku 4.5 at 0.0%

Anthropic still says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those on Terminal-Bench 4.0. Haiku 5.5 is framed for narrowly scoped work that used to be too expensive to run often: compaction, summarization, and subagent calls.

Treat those scores as Anthropic's evaluations. They are directionally useful. They are not an independent warranty for your CRM or inbox.

Customer quotes Anthropic published

Anthropic also published early customer comments. These are vendor-relayed, not independently audited:

• Asana: "over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn"
• HubSpot: 92.8% averaged over three runs on a CRM evaluation suite
• AlphaSense: about 8M Ask in Document calls a week, with Haiku 5.5 scoring 0.84 versus 0.76 for Haiku 4.5 on 400 queries

Useful as signals that high-volume document and CRM work is where Anthropic wants Haiku 5.5. Not useful as a substitute for your own eval set.

What this is not

Haiku 5.5 is not a reason to put your hardest judgment calls on the cheapest model. Anthropic's own framing keeps Sonnet and Opus for complex agentic coding. Cheap tokens still invent fields, mis-tag tickets, and miss edge cases if you remove the review step.

It is also not proof that every subagent pattern is production-ready. Speed and price change how often you can afford to call a model. They do not remove ownership of exceptions.

How to evaluate it this week

1. Pick one high-volume path you already measure: inbox triage, CRM lookup, document Q&A, or a coding subagent that pulls a single fact.
2. Log cost per successful task, latency, and error rate against Haiku 4.5 or whatever you use today.
3. Keep Sonnet or Opus on drafting, multi-step plans, and anything customer-facing without a gate.
4. Recalculate after retries and human cleanup. Token price alone lies.

What it means for operators (Jaden's take)

I care about the model that runs a hundred times a day more than the model that wins a keynote. Haiku 5.5 is interesting because Anthropic put the price under a dollar per million input tokens for the common request size and named the exact jobs operators already burn tokens on: classification, summaries, database queries, and subagents.

My rule is simple. Put the cheap, fast model on work that is easy to verify. Keep the expensive model, or a person, where a wrong answer is expensive. If Haiku 5.5 can pull a CRM field or classify a ticket with a spot check, that is leverage. If it quietly misroutes revenue work and you only notice in a weekly review, that is debt.

Re-price your high-volume paths this month. Keep Sonnet and Opus for the hard steps. Ignore the average "75% cheaper" line until your own logs agree.