Claude Haiku 5.5: when the cheap model is enough

On 7 October Anthropic released Claude Haiku 5.5. The headline is speed and price, not a new kind of model. They say it costs around 75% less than Haiku 4.5, and that it is their fastest Haiku yet. For a small team that already routes work across models, that is a day to run evals, not a day to rewrite the product.

We run one backend for Cabin and for Mo inside QuesMo. Some turns need a short, honest opinion. Some turns only need a title, a summary or a label. This post is how we would try Haiku 5.5 on that backend, and what we would refuse to move until the numbers say so.

Key takeaways

• Haiku 5.5 is for volume: summaries, labels, compaction, narrow subagent jobs.
• Anthropic prices prompts up to 100k tokens at $0.10 input and $0.50 output per million tokens.
• We would trial it on the cheap tasks first, not on Cabin’s actual reply.
• Sonnet 5.5 cache reads are now half the old price, which helps the work we already cache.
• No model change reaches users here until it beats the current setup on offline evals.

What actually got cheaper

For the short prompts that make up most Haiku traffic, Anthropic’s list is:

• Input: $0.10 per million tokens (was $1.00 on Haiku 4.5)
• Output: $0.50 (was $5.00)
• Cache reads: $0.01 (was $0.10)
• Cache writes: $0.125 (was $1.25)

Above 100,000 input tokens the new prices are higher, but still half of Haiku 4.5. Anthropic says about 90% of Haiku 4.5 requests sat under that line, which is why they quote “around 75% less” as the average.

The same day, cache reads on Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens. They estimate that cuts most agentic Sonnet work by about 20%. If your prompts are laid out so the stable part can be cached, that cut lands without any other change.

What we would move

In how we route between providers, the cheap path already exists. Titles, conversation summaries, memory labels, and “what kind of message is this” are not the product. They have to be good enough. They do not have to be the best model in the catalogue.

Those are the first Haiku 5.5 candidates:

• Titles and summaries. High volume, easy to score (length, no empty output, no leaked system text).
• Classification. Intent, language, whether the person asked to delete memory.
• Compaction. The short recap that keeps a long thread inside the context window.
• Narrow subagents. Pull one fact from a long note while a larger model writes the reply.

Anthropic’s own customers describe the same split. Rogo uses Haiku as a subagent to pull a line from a 10-K while a bigger model builds the deck. That is the shape we already want: a small model for the fetch, a larger one for the judgement.

What we would not move yet

Cabin’s reply is the product. Two or three lines, an actual opinion, in the language the person used, including Hinglish. A cheaper model that hedges, lectures, or slips into English on a Hindi turn is a worse product, even if the invoice looks better.

So the main Cabin turn, and Mo’s chat turn, stay on the current model until an eval set of real conversations says Haiku 5.5 is as good or better on:

• reply length
• whether it takes a side
• whether it matches the user’s language
• whether it invents memory

We would rather spend a day measuring that than a week undoing a noisy launch. The method is in LLM evals for small teams.

Complex agentic coding is the other thing we would not send here. Anthropic is blunt about it: Sonnet and Opus still win on Terminal-Bench. Haiku 5.5 is not a replacement for the lead model in a software factory. It is a better worker for the short jobs around that lead.

How we would actually switch

Routing here is a setting, not a deploy. The trial would look like this:

1. Point only the title and summary tasks at claude-haiku-5-5.
2. Replay the offline set for Cabin and for Mo.
3. If both pass, ship that slice. Leave the main reply alone.
4. Watch latency and cost for a few days. Then consider classification and compaction.

One slice at a time, with a way back. That is the whole trick. A new model launch is a bad day to change everything.

The Sonnet cache-read cut can go live even sooner, because it does not change the model. Check that your prompt still puts the stable block first. If it does, you are already paying less for the same work.

The bottom line

Haiku 5.5 is a cheaper, faster small model, not a reason to collapse every route onto one name. Use it where you were already overpaying. Keep the turns people notice on a model you have measured. Run the evals you should have had before the launch email arrived.

We build this kind of routing and gating into client work as LLM app development. If you want a second pair of eyes on a model switch, write to us.

FAQ

What is Claude Haiku 5.5?

Anthropic's small, fast Claude model, released on 7 October 2026. It is meant for high-volume work such as summaries, classification and subagent tasks, not for the hardest coding or the most careful replies.

How much cheaper is Haiku 5.5 than Haiku 4.5?

Anthropic says it costs around 75% less on average. For prompts up to 100,000 tokens, input is $0.10 per million tokens and output is $0.50, against $1.00 and $5.00 on Haiku 4.5.

Should I send every request to Haiku 5.5?

No. Use it where a cheaper model was already good enough, or where a subagent does a narrow job. Keep the main reply, the hard decision, and long agentic coding on a larger model until your evals say otherwise.

What else changed in the Claude price list this week?

Cache reads on Claude Sonnet 5.5 were cut in half, to $0.10 per million tokens. Anthropic says that makes most agentic Sonnet work about 20% cheaper.