Claude Haiku 5.5 in BrainBox: benchmarks, pricing and when to use it

Claude Haiku 5.5 has been selectable in the BrainBox Assistant since October 7, 2026, the day Anthropic released it. It appears in the Default category of the picker with a one million token context window, 128,000 max output tokens and image input. It succeeds Claude 4.5 Haiku, which stays available, and in BrainBox uses about a third of its intelligence units on the same task.

What is Claude Haiku 5.5 and what improved over Claude 4.5 Haiku?

It is Anthropic's small model for high-volume work. Anthropic calls it "the cheapest, fastest, and most capable small model we've ever released" and recommends it for summaries, classification, database queries, live customer support and as a subagent for Opus 5.5. The official model page lists it as the fastest in the family, with adaptive thinking and a June 2026 knowledge cutoff.

It is also the first Haiku with an adjustable effort level: more effort means more reasoning and more usage. In Anthropic's own benchmarks the jump over Haiku 4.5 is large:

The best score in each row is in bold.

Anthropic includes Sonnet 5.5 as a reference, and it wins every row, as you would expect from a larger model. Sonnet 5.5 is not in the BrainBox picker today. The comparison that matters here is GPT-6 Luna, which costs the same on short questions, and Haiku 5.5 beats it on every row with data. Luna's numbers are reported by Anthropic, not OpenAI.

Independent measurement points the same way. Artificial Analysis gives it 43 on its Intelligence Index at max effort, 38 at high effort and 34 at medium effort, against a median of 13 in its price range. GPT-6 Luna scores 38 and Claude Opus 5.5 scores 58.

How does the BrainBox Assistant work with Haiku 5.5?

The BrainBox Assistant is an agent. You choose which Boxes and files it can use and which model does the reasoning; the agent decides the steps. Each one shows up in the chat: "Listing sources", "Searching documents", "Reading document" with the pages it opened, "Saving note" when it sets something aside, and finally the answer with page citations.

Most of what the agent does is short, repeated steps: search, read a few pages, summarize, cite. Haiku 5.5 is built for that. Box reports that in early testing it scored 11 points higher than Haiku 4.5 at about half the latency.

1. Open the Assistant and pick your sources

   In the sidebar you check the Boxes and files the agent can use. You can also point to a file with @ inside the message.

2. Click the model picker

   It sits in the bottom bar of the chat. A searchable list opens.

3. Type 'Haiku'

   Claude Haiku 5.5 and Claude 4.5 Haiku appear, with their context window and the Images label.

4. Choose Claude Haiku 5.5

   If you cannot see it, the workspace model policy does not allow it.

5. Ask for the task

   Page citations, tools and the agent's visible steps work the same with any model.

Select model

Select model

Haiku

• 

  Claude Haiku 5.5Anthropic

  Fastest, cheapest Claude for high-volume work

  1M contextImages
• 

  Claude 4.5 HaikuAnthropic

  Previous Haiku, fast with near-frontier intelligence

  200K contextImages

BrainBox is an AI workspace for documents that answers with citations to the exact page. A smaller model reasons in fewer steps, and every citation still leads to its page.

How many intelligence units does Haiku 5.5 use per task?

In BrainBox, usage is measured in intelligence units, not tokens, and each task uses a different amount depending on how many searches and page reads the agent runs, how much the model reasons and how much it writes. Estimates for the standard plan:

These are estimates: a longer answer, more sources or more agent steps push the number up. On short tasks Haiku 5.5 and Luna use about the same; on tasks that pile up many pages read, Haiku 5.5 climbs a bit more. For reference, the Elite plan includes 400 units a month, enough for about a hundred case file reviews with Haiku 5.5.

Which tasks is it worth choosing for?

A compliance team can keep Haiku 5.5 as the everyday model and enable Opus 5.5 only for final reviews, from the workspace model policy. Citations point to the page with either one, so verifying the answer works the same.

What about safety?

Anthropic reports that Haiku 5.5 improves on Haiku 4.5 across almost all of its alignment evaluations, with far fewer cases of misaligned behavior and less willingness to cooperate with misuse. These are the company's own evaluations, described in its system card. In the Assistant, the practical protection is the same as before: the agent asks for approval before sensitive actions on your files, and every claim carries its citation.

HubSpot says Haiku 5.5 scored 92.8% on its internal test suite, averaged over three runs, the best result it has seen on that suite.

Sources

• Anthropic, Introducing Claude Haiku 5.5: October 7, 2026 announcement, benchmark table, cost against Haiku 4.5, effort level, alignment and customer quotes.
• Anthropic, Claude Haiku 5.5 model page: one million token context, max output, adaptive thinking, modalities and knowledge cutoff.
• Anthropic, Claude Haiku 5.5 system card: methodology of the safety evaluations.
• Artificial Analysis, Claude Haiku 5.5 pages at max, high and medium effort: Intelligence Index, price-range median, output tokens and verbosity.
• Artificial Analysis, GPT-6 Luna page: reference score on the same index.