Claude Sonnet vs Opus: How to Route Agent Work to Cut Costs

If you're running the biggest, smartest Claude model for every task your agent touches, you're overpaying by a wide margin and probably not getting better results for it. Most agent work is not hard reasoning. It's coding edits, browsing, data cleanup, and routine tool calls that a mid-tier model handles just as well for a fraction of the price.

Quick answer: Claude Sonnet vs Opus comes down to task difficulty, not defaulting to the strongest model. Use Sonnet as your everyday default for coding, browser, and agent work; escalate only genuinely hard reasoning to Opus; and route simple, high-volume, latency-sensitive tasks to Haiku. Layer prompt caching (up to 90% off repeated input) and the Batch API (a flat 50% off) on top and most agent bills drop sharply with no quality loss.

Which Claude model should agents use by default?

Sonnet is the right default for the overwhelming majority of agent work: writing and editing code, driving a browser, calling tools, drafting research, updating records, and running multi-step workflows. It's fast enough to feel responsive in a loop and cheap enough to run all day. Anthropic launched Sonnet 5 at introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026 (then $3/$15), which improves the economics of text-heavy agent workflows. Actual cost still depends on your prompt size, tool calls, and retries.

The mistake founders make is treating the most capable model as the safe choice. It isn't safe on a budget. The moment your agent runs in a loop — read a ticket, look something up, draft a reply, check it — the biggest model multiplies a cost you didn't need to pay. A ten-step agent isn't ten times one call; each step carries the growing context of everything before it, so the input bill compounds fast. Run that loop on Opus for tasks Sonnet handles fine and you've turned a modest per-call gap into a monthly surprise.

When is Opus worth the extra cost?

Opus earns its price on the small slice of work where a wrong answer is expensive and the reasoning is genuinely hard: multi-file architecture decisions, gnarly debugging across a large codebase, dense legal or financial analysis, ambiguous planning where the agent has to weigh tradeoffs it can't just look up. These are the tasks where Sonnet stalls or produces something you'd have to redo, and redoing work is the most expensive thing an agent does.

The way to use Opus well is to escalate to it, not start on it. Let Sonnet attempt the task. If it flags low confidence, produces something that fails your verification step, or hits a problem that needs deeper planning, hand that specific step to Opus. This mirrors how a good team works — the senior person gets pulled in for the hard call, not every routine one. Coursiv, which reviews AI cost strategies, calls model routing the single most effective lever for controlling spend, and this is why: you're paying premium rates only for premium-difficulty work.

When should you route work to Haiku?

Haiku is the right choice for high-volume, latency-critical, low-complexity tasks where speed and price matter more than depth: classifying incoming emails, tagging and routing support tickets, extracting fields from a form, simple yes/no checks, or the first-pass triage that decides where a request goes next. If a task is something a sharp intern could do in five seconds without thinking hard, it's a Haiku task.

The pattern that works is a cheap model doing the sorting and a stronger model doing the thinking. In a support agent, Haiku can read every inbound message and decide whether it's a password reset (handle instantly), a billing question (pull the account), or an angry escalation (route to a human). Only the messages that need real work ever reach Sonnet. We build support agents this way because it keeps the per-ticket cost low without dropping resolution quality — the support agent that resolves tickets instead of deflecting leans on exactly this split.

A simple routing rule you can apply today

Sort every task your agent does by difficulty and cost sensitivity, then assign a tier — the same way the automate, augment, or strategize sorting works for deciding what to hand off in the first place. Here's the default mapping most agent builders can start with:

Keep this version-agnostic. The names and prices change; the rule doesn't. Default to the mid-tier model, escalate difficulty up and route volume down. When a new model generation ships, you re-slot the same three tiers rather than rethinking your whole system.

The two cost levers bigger than model choice

Routing gets you most of the way, but two Anthropic features cut input cost harder than any model swap, and most builders leave them off. Prompt caching stores the stable part of your prompt — a long system instruction, a reference document, a tool schema — so you don't pay full price to send it on every call. Anthropic prices cached input reads at up to 90% off the base input rate. For an agent that reuses the same 5,000-token instruction on every one of a thousand daily calls, that's the difference between a real bill and a rounding error.

The Batch API takes a flat 50% off both input and output for work that doesn't need an instant answer. Anything you can queue — overnight research reports, bulk CRM enrichment, processing a backlog of documents, generating a hundred summaries — belongs here. You submit the jobs, Anthropic returns them within 24 hours, and you pay half.

Stack all three moves and the savings compound: Sonnet instead of Opus as your default, caching on the repeated context, batching everything that isn't live. That's how a workflow that looked expensive on the biggest model becomes cheap enough to run at scale.

Manual overspending vs a routed setup

The default way people run agents is one model for everything, usually the strongest, because it feels safer and it's less to think about. A routed setup costs a little upfront thought and saves continuously after that. The contrast is stark once you put real numbers on a loop.

The one-model habit isn't wrong because the model is bad — it's wrong because you're paying premium reasoning rates for triage and data entry. That gap is pure waste, and it grows every time the agent runs.

Common pitfalls when routing model work

The most common mistake is over-escalating: routing anything that feels important to Opus. Importance isn't difficulty. A high-stakes email still just needs a competent draft, and Sonnet writes it fine — reserve Opus for tasks that are genuinely hard to reason through, not merely tasks that matter.

A few others that quietly cost money:

FAQ

Is Claude Sonnet or Opus better for coding agents?

Sonnet is the better default for coding agents. It handles most edits, refactors, and tool calls quickly and cheaply, and it runs well in the tight loops agents depend on. Escalate to Opus only for genuinely hard problems like multi-file architecture decisions or deep debugging across a large codebase, where a wrong answer costs more than the extra spend.

How much can model routing actually save on Claude costs?

Model routing plus caching and batching can cut a Claude agent bill by a large margin because you stop paying premium rates for routine work. Prompt caching alone reaches up to 90% off repeated input, the Batch API takes a flat 50% off non-urgent jobs, and moving default work from Opus to Sonnet lowers the per-call rate on top of both.

Which Claude model is cheapest for high-volume tasks?

Haiku is the cheapest Claude model and the right choice for high-volume, latency-critical tasks like classification, tagging, field extraction, and first-pass triage. Pair it with Sonnet so the cheap model sorts and routes while the stronger model only handles the requests that actually need it.

Should I always use the newest, most capable Claude model?

No — defaulting to the most capable model for everything is the most common way founders overspend on agents. Match the model to the task: mid-tier for everyday work, the top model only for hard reasoning, and the smallest model for high-volume simple tasks, re-slotting the same three tiers whenever a new generation ships.

Routing is a skill you build once and reuse across every agent you deploy, and it's exactly the kind of setup we walk through with builders inside the free Claude Community — join in, share your workflow, and get your tiers dialed in with people running the same playbook.