Unlocking Haiku 5.5 and Effort Levels in Claude Code

Lachlan Reed

Brought to you by Jellypod AI. Last episode I said asking for max effort on a Haiku subagent would do nothing, because Haiku had no effort levels. That caveat lasted about a day. Claude Code 2.1.293 shipped on October 7th with Claude Haiku 5.5, the first Haiku with an adjustable effort dial: low, medium, high, xhigh, and max, defaulting to medium.

James Turner

The changelog line says Haiku 5.5 is now the default Haiku model on the Anthropic API, so the haiku alias resolves to it. You get a 1M token context window, up from 200K. Pricing is ten cents per million input tokens and fifty cents per million output tokens for prompts up to 100,000 tokens. Over 100,000, it jumps five times, to fifty cents and two fifty. For comparison, Haiku 4.5 was a dollar in and five out, and Sonnet 5.5 input is two dollars. So it's 90 percent cheaper per token than Haiku 4.5, and 20 times cheaper on input than Sonnet.

Lachlan Reed

Cheap used to mean weak, so here's the capability side. These are Anthropic's own benchmarks, not independently tested. On Terminal Bench 4.0, command line coding tasks, Haiku 4.5 scored zero percent and Haiku 5.5 scores 39.2. VentureBeat reads that chart as roughly 39 at max effort and about 20 at the medium default. Sonnet 5.5 scores 70.6, so this is not your lead model. Anthropic itself says Sonnet and Opus stay better for complex agentic coding, and Haiku fits narrow jobs like compaction, summarization, and subagent work. Rogo's example: while a bigger model builds the deck, a Haiku subagent pulls one revenue line out of the annual filing.

James Turner

Here's how to use it. In a session, run slash model haiku. But the real win is subagents, and there's a catch: the built in Explore subagent runs on your main model, so it won't get cheaper by itself. Define your own. Make a file in dot claude slash agents called scout dot md, with frontmatter: name scout, a description like searches the repo and summarizes what it finds, and model haiku. Now every search that scout runs bills at Haiku rates.

Lachlan Reed

Then pair it with the effort parameter on the Agent tool from 2.1.292. Prompt something like: use the scout subagent at low effort to list every caller of parseConfig. Later: use scout at max effort to work out why the config test fails. Low for mechanical searching, max when the question is narrow but tricky. Same cheap worker, different thinking budget.

James Turner

Now the caveats. First, the 100,000 token cliff. A 99,000 token prompt costs about a cent to send. A 101,000 token prompt costs about five cents. So keep subagent prompts scoped. Second, the tokenizer is new and counts more tokens for the same text. Simon Willison measured about 1.25 times on one long prompt, and one video breakdown said around 30 percent, so part of that 90 percent discount gets eaten. Anthropic's own average is about 75 percent cheaper per job. Third, the alias only points to Haiku 5.5 on the Anthropic API. On Bedrock, Google Cloud, and Microsoft Foundry, haiku still means Haiku 4.5.

James Turner

Fourth, and this matters most: Anthropic's guide says that at low and medium effort, Haiku 5.5 sometimes reports a code change as done without running a check. So keep your tests or a verifier hook in the loop. And if you call the API directly, custom temperature and top p, thinking budgets, and prefilled replies return errors.

Lachlan Reed

A few smaller fixes. In 2.1.294, prompt and agent hooks written as plain instructions, like block commands that do some dangerous thing, were actually allowing what they should block. That's fixed. Stop and SubagentStop hooks like carry on if the build is broken are also judged better, so Claude is less likely to quit early. If you write hooks in plain English, update today.

James Turner

And in 2.1.293, path scoped rules and nested CLAUDE.md files now load when Claude reads a file with a single file cat, head, tail, sed with the n flag, or grep in Bash, not only through the Read tool. Also fixed: Claude treating its last actions before a compaction as finished afterward and redoing work, plus a memory leak where an HTTP MCP connection kept every request it had sent. And subagentStatusLine now includes agentType, so your status scripts can tell custom subagents apart.