The Subagent "Grunt Work" Just Got 10x Cheaper — Claude Haiku 5.5 Leaps from Zero to 39% on Terminal-Bench and Becomes Claude Code's Default Haiku
Anthropic has released Claude Haiku 5.5, a new version of its small model. Priced at roughly one-tenth of the previous generation and expanded to a 1M-token context, it jumps from 0% to 39.2% on the agentic coding benchmark Terminal-Bench. Adopted as Claude Code's default Haiku, it settles into the
On October 7, 2026, Anthropic released a new version of its small model, "Claude Haiku 5.5." Alongside it, Claude Code v2.1.293 switched its default Haiku to this model (model ID claude-haiku-5-5). Haiku keeps its position as small, fast, and low-cost, but the headline this time is a major jump in its capability on agentic tasks.
A Price Tag Approaching One-Tenth
Pricing is split into two tiers by prompt length. In the standard tier, up to 100K tokens, the price dropped sharply from the previous-generation Haiku 4.5's "$1 input / $5 output." The context limit expanded from 200K tokens to 1M tokens, and the maximum output is 128K tokens.
| Prompt length | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Up to 100K tokens | $0.10 | $0.50 |
| Over 100K tokens | $0.50 | $2.50 |
Cache reads are $0.01 / $0.05, and cache writes are $0.125 / $0.625 (standard tier / extended tier, respectively). According to Anthropic, it is on average about 75% cheaper than Haiku 4.5, and for use cases under 100K tokens the savings reach up to 90%.
Benchmarks: A Leap from Zero
What stands out is the gains on agentic tasks, which small models have historically struggled with. Here are the numbers compared with the previous generation.
- Terminal-Bench 4.0 (agentic coding): 39.2% (Haiku 4.5 scored 0%)
- OSWorld 2.1 (computer use): 72.4% (vs. 15.7%)
- GDPval-AA v2.1 (knowledge): 1,620 (vs. 735)
Its ability to get work done by typing commands in a terminal went from a level where the previous generation was effectively unusable to a range where it can handle a certain amount of real work. "Computer use"—looking at a screen and operating it—also improved dramatically.
Where It Pays Off — Not the Lead, but the "Grunt Work"
Anthropic positions this model as best suited for "narrowly scoped tasks." Concretely, that means uses like the following.
- Compaction (summarizing and compressing long conversations)
- Summarizing search results and logs
- Running subagents spun off by a parent agent
At the same time, Anthropic is explicit that Sonnet 5.5 and Opus 5.5 are better suited for complex agentic coding itself. In other words, Haiku 5.5 is not a model meant to take the lead in reasoning; its role is to handle, cheaply and quickly, the "subcontracted work" that the lead model throws at it in bulk.
Impact on Workflows and Business
This character meshes with the "one director plus multiple subagents" structure that various tools are now pushing toward. In the immediately preceding v2.1.292, Claude Code added an effort parameter for specifying how much effort a subagent puts into its work. This makes it easier to assign Haiku 5.5 to light grunt work at low cost while routing only intricate design and fixes to higher-tier models.
From a management standpoint, the cost structure of running agents around the clock changes. When the unit price of repetitive work—summarizing, classifying, preprocessing—falls, it becomes easier to run automations that were previously shelved as not worth the cost. The 1M-token context is also well suited to preprocessing that reads a large repository or logs all at once. It is available on the Anthropic API as well as AWS, Google Cloud, and Microsoft Azure.
The Pitfall That Cheapness Invites
A lower unit price does not directly mean a lower total bill. The cheaper each call is, the more freely teams tend to add loops and subagents, and as a result the number of calls can balloon—potentially driving the total spend and compute consumption higher instead. For unattended operations that run nonstop, the safe move is to put cost visibility and spending caps in place first.
On quality, too, Anthropic itself draws a line, saying higher-tier models are better for complex coding. Handing heavy judgment calls to a grunt-work model just because it is cheap can increase rework and end up costing more in some cases. Choosing which model to use is best decided by task difficulty, not the price sheet alone.
References: Claude Code changelog (v2.1.293) / The Decoder: Claude Haiku 5.5 / MarkTechPost: Claude Haiku 5.5 / LLM-Stats: Claude Haiku 5.5
