Claude Sonnet 5.5 & Haiku 5.5: New Anthropic Models in Smoke Monkey

Anthropic's October 2026 releases split the job cleanly. Claude Sonnet 5.5 raises the ceiling for agentic coding — multi-file changes, long plans, and tool use — while Claude Haiku 5.5 drives the cost curve down for mechanical subtasks like summarization, classification, and result filtering. The practical move is not to pick one; it is to pair them: a Sonnet 5.5 planner coordinating Haiku 5.5 workers. But a stronger model alone does not give you scheduling, tools, permissions, or observability. That still comes from the runtime, which is where Smoke Monkey Harness and Smoke Monkey Canvas come in.

Because Smoke Monkey Harness treats every provider as an adapter, adopting Sonnet 5.5 or Haiku 5.5 does not touch your tools, prompts, or permission model. You change the model string and keep the 6-phase loop, 24 built-in tools, and permission gates exactly as they were. That is the difference between a vendor SDK — where each new model is a migration — and a provider-neutral runtime, where the model is a single line. Assign Sonnet 5.5 and Haiku 5.5 to different cards on Smoke Monkey Canvas and compare spend live.

The real payoff of having Sonnet 5.5 and Haiku 5.5 in one runtime is tier routing. Send high-judgment work — planning, security review, tricky refactors — to Sonnet 5.5, and route high-volume, low-judgment work to Haiku 5.5. Because the harness is model-agnostic, this is a per-agent setting, not a rewrite, and it stacks with token cost optimization and context compaction. You get frontier quality where it matters and cheap throughput everywhere else. On Smoke Monkey Canvas each tier is a labeled card, so routing is visible at a glance.

Run npx @smoke-monkey/canvas start and your Claude-powered fleet becomes a living board. Assign Sonnet 5.5 agents to planning and review cards, Haiku 5.5 agents to execution cards, and watch tool calls stream in with per-agent token usage. Approve destructive steps inline through human-in-the-loop gates, and because Canvas is built on the same open runtime, you can mix in a local or alternative provider on any card without leaving the workspace. It is the fastest way to see exactly what the newest Anthropic models do with your repository. Related: multi-provider agent APIs and Claude Opus 5 multi-agent swarm. It all runs on Smoke Monkey Harness, which keeps the models swappable and the loop identical.