Tiered Routing

Tiered routing is the 2026 default architecture for any production agent built on Claude. Instead of choosing one model for the entire product you build a thin decision layer that inspects each incoming job and routes it to Opus 4.7, Sonnet 4.7, or Haiku 4.7 depending on what the step actually needs. Opus gets the long horizon reasoning that must stay coherent for two to four hours, the hard ranking problems, the complex tool orchestration that involves nested decisions. Sonnet gets the daily driver work, the multi file code refactors, the structured data extraction, the main body of any agent loop. Haiku gets the high frequency surface work, the initial classifications, the real time streaming responses, the quick lookups and light drafting. The magic is that 4.7 made all three variants share the same 1M context window, the same rock solid JSON tool use that stays above 95 percent reliability, the same computer use primitives, and the same prompt caching options including the new one hour tier. That uniformity means the router can hand context back and forth without translation layers or lossy summarization steps. In practice the router is often just a system prompt that asks a cheap model to classify the task difficulty on a 1 to 5 scale plus a few rules based on keywords from your domain. Add prompt caching and the unit economics become unbeatable. A customer support agent that loads a 200K token knowledge base pays full price once then enjoys cheap cached reads on Haiku for the rest of the session. The article on Claude 4.7 for builders calls this the cheapest production stack and it is. Teams no longer face the 2024 dilemma of smart versus fast. They get smart, fast, and cheap in the right combination. The article provides a clear capability split. Opus for long horizon agents, hard ranking, planning, complex tool orchestration with the slowest latency and highest cost. Sonnet for daily driver agents, code editing, multi file refactors with mid latency and the workhorse cost. Haiku for high throughput, voice, chat, light drafting with the fastest speed and cheapest price. The rule of thumb is Opus for the hard reasoning step, Sonnet for the agent loop body, Haiku for the high frequency surface. Teams that follow this split report the best quality to cost ratios.

Tiered routing is not just using the cheapest model that might work. It is not a set it and forget it config you tweak once then never revisit. It is not microservices hell with separate queues for each model. It is not an excuse to ship sloppy prompts because the router will save you. It is not the right choice for every single use case. Builders who implement naive routers based only on message length instead of semantic difficulty end up with agents that feel erratic. The router itself needs testing and iteration just like the rest of the system. Treat it as core product logic not infrastructure trivia.

The concrete shipping examples in 2026 all follow the same script. Linear AI uses tiered routing for their issue lifecycle. Every new issue starts on Haiku for rapid labeling, urgency detection, and linking to similar past tickets. The router looks at the issue description length, the presence of words like blocker or redesign, and the number of linked PRs. Anything scoring above a threshold or containing planning language routes to Sonnet for comment drafting and suggested next steps. The final prioritization that weighs business impact against engineering effort hits Opus. This keeps their average cost per issue low while the quality of the AI assistance feels consistently excellent. Devin pushes the pattern to the limit with overnight autonomous coding tasks. File system navigation, simple bug fixes, and test generation run on Haiku. The core loop of understanding the ticket, planning the changes, and executing multi file edits runs on Sonnet. When the agent reaches a point where previous approaches have failed or when it needs to propose a brand new algorithm it escalates to Opus for fresh eyes. The stability gains from 4.7 combined with this routing let Devin complete real tickets end to end. Cursor routes tab completion on Haiku for the sub 100ms responses developers expect then hands anything wider than a single function to Sonnet. Granola runs live transcription on Haiku but promotes the meeting summarization and action item parsing to Sonnet because that step requires more contextual understanding. A design critique tool built by a San Francisco agency routes pixel level audits and accessibility scans to Haiku, component level redesigns that respect the design system to Sonnet, and full product strategy critiques that incorporate market positioning to Opus. One rubric based eval pipeline at a design systems team routes initial output generation to Sonnet but sends the actual fifteen criterion grading and detailed explanation of deductions to Opus because consistent rubric application demands the heavier model. The batch API makes this cheap enough to run on every prompt change. These examples all share one trait. The router was tuned by logging thousands of real interactions then adjusting the thresholds until the escalation rate matched the actual difficulty curve.

Use tiered routing as soon as your agent crosses the threshold from prototype to production or as soon as you see your model bill climbing faster than your user growth. The Claude 4.7 release removed every technical objection. Long horizon stability past two hours, family wide 1M context, twice as fast computer use, JSON modes that no longer need retry wrappers, Sonnet forty percent faster, and real time Haiku made the tiered approach not only possible but obviously correct. Use it for anything that mixes quick surface interactions with deep backend reasoning. Use it for full codebase editors that hold seventy five thousand lines in context and still make smart architectural choices. Use it for computer use flows doing browser automation and QA where latency on every click actually matters. Skip it during the first two weeks of exploring a new idea. A single model keeps the prompt surface smaller and the debugging easier. Skip it also if every single call in your system requires the same level of intelligence. Then just pick the model that matches that level and call it a day. Most teams however find that careful measurement reveals 60 percent of turns could run on Haiku, 35 percent need Sonnet, and only 5 percent truly need Opus. Capturing those percentages is what makes tiered routing the highest leverage change most AI builders can make in 2026.

Tiered routing turns the Claude 4.7 family from three separate models into one intelligent unified system that assigns exactly the right intelligence to every single task.