Haiku 5.5 is live in both the Claude Platform and Claude Code at approximately 75% lower cost than Haiku 4.5. It is designed to work as a subagent alongside Opus 5.5 or Sonnet 5.5 for high-volume, cost-sensitive tasks such as summaries, compactions, and database queries.
Haiku 5.5 is the newest member of the Claude 5.5 family and the first Haiku to support effort control. Pricing is $0.10/$0.50 per million tokens for prompts up to 100k tokens, rising to $0.50/$2.50 above that threshold.
The tiered pricing structure for Haiku 5.5 is $0.10 input / $0.50 output per million tokens for prompts under 100k tokens, with cache reads at $0.01. Above 100k tokens it rises to $0.50 / $2.50 with cache reads at $0.05.
Haiku 5.5 shows notably better prompt injection resistance than Haiku 4.5, landing slightly above Gemini's Astra/Luna but a bit below Gemini Flash according to internal benchmark data (page 50 of the cited report).
Starting in v2.1.292+, you can instruct Claude to run subagents at a specific effort level, giving finer control over how much compute each subtask consumes.
Before version 2.1.292, subagents simply matched the session's effort level; asking Claude to run them at a different effort had no real effect. The new release adds an effort parameter to the Agent tool, so verbal instructions to adjust effort now actually work.
Desktop support for the new subagent effort-level feature is coming very soon, as a follow-up to the v2.1.292+ release that introduced the ability to specify effort levels when running subagents.
The architectural direction for Claude Code is a split model: reasoning and 'brains' live in the cloud, while a lightweight local agent provides hands-on access to your machine and files.
Autocompact thresholds are stored per model, so setting it to 100k on Haiku leaves other models untouched. You can also configure it directly in ~/.claude/settings.json under modelSettings for 'claude-haiku-5-5'.
The Python and TypeScript SDKs now include computer use and browser use toolsets natively. Previously developers had to write their own action loops; the SDKs now handle the loop and send actions to drivers automatically.
Compatible drivers for the new built-in computer use and browser use toolsets include Browser Use, Browserbase, E2B, and Daytona, or you can write your own following the quickstart example drivers.
A user reported that their entire C drive was deleted, reportedly by Opus 5.5, and they were only saved by NAS backups. The Claude Code team responded that auto mode for permissions almost certainly would have caught the action and asked why it wasn't enabled.
A user called auto mode 'beyond annoying' and said it creates more blockers than it solves. The Claude Code team responded asking for specifics in order to improve it.
Users report that auto mode blocks innocuous commands mid-workflow, forcing them to drop to 'accept edits' or 'manual' mode just to surface the approval prompt, then switch back. The Claude Code team is actively soliciting specific examples to fix the behavior.
A user reports that auto mode repeatedly overrides their authorization for the same action, requiring them to disable it before the action can proceed. The Claude Code team asked for a /bug report.
Monthly API credits are rolling out: Max 5x gets $100, Max 20x gets $200, and Team plan receives up to $500 pooled across the team. Credits work on any model, including Haiku 5.5, in direct API calls or third-party tooling.
