The right answer to "which Claude model should I use for my agent?" is all three — just not for the same job. Put Haiku 4.5 behind the mechanical, high-volume steps (classification, extraction, query rewriting); make Sonnet 5 your workhorse default for reasoning, code, and tool loops; and reserve Opus 5 for the hardest, rarest calls where a wrong answer is expensive. One caveat with a date on it: Sonnet 5 is running a temporary promo of $2 / $10 per million tokens that ends August 31, 2026, after which it returns to $3 / $15 — a real, dated cliff that changes the routing math. (All prices below are per million input/output tokens, verified against Anthropic's pricing page as of August 2026 — verify current pricing before you commit, because these move.)

When to reach for each#

An agent is not one model call. It's a pipeline of them — some trivial, some hard, most run over and over. Price each job, not the agent, and the tiers fall out naturally.

Haiku 4.5 — $1 / $5. The high-volume tier. This is the cheapest per token, and most agent calls are mechanical: deciding which branch to take, pulling structured arguments out of a message, rewriting a user question into a retrieval query, drafting a throwaway first pass, summarizing at scale. None of these reward a bigger model — the work is well-specified and low-stakes. Haiku is where the 80% of calls that are plumbing belong. Because it's a fifth the input price of Opus 5, moving this volume off the expensive tier is where most of your savings actually come from.

Sonnet 5 — $2 / $10 today, $3 / $15 on Sept 1. The workhorse default. When a step needs genuine multi-step reasoning, writes code, or drives a tool-use loop where a bad tool call cascades into more bad calls, Sonnet 5 is the balance point — near-Opus quality on coding and agentic work at a fraction of the cost. Make it your default and only move off it deliberately. The one thing to keep in mind: it gets 50% more expensive on output in September, which narrows its gap to Opus and is the whole reason this piece has a deadline.

Opus 5 — $5 / $25. The escalation tier. Opus 5 is for the hardest and least frequent calls: planning a long-horizon task, untangling a thorny bug, and the final review or verification step where shipping a wrong answer costs real money. Used as an escalation — triggered by a confidence threshold or a specific step type, not as a default — Opus barely registers on the bill, because it runs on a small slice of calls while carrying the expensive ones. The mistake is defaulting to it "to be safe." That's how a $30 agent becomes a $90 one.

Fable 5 — $10 / $50. The specialist. Anthropic's most capable widely released model, for the most demanding long-horizon reasoning. For most founder agents it's overkill; reach for it only when you can show Opus 5 measurably isn't enough. Mentioned here so you know the ceiling exists — not so you build on it.

A reference agent, tiered#

Make it concrete with a customer-support agent handling, say, 1,000 tickets. Every ticket flows through the same pipeline, but each step runs on a different tier:

Here's the rough shape of the bill, per 1,000 tickets (illustrative token counts, not measured — your mileage will vary):

That's roughly $28 per 1,000 tickets. Put Opus 5 behind every step instead — classification, drafting, and all — and the same volume lands near $90. The tiered version is about 3x cheaper, and it's faster too, because Haiku and Sonnet return quicker on the calls that dominate the count. Layer on prompt caching (cached input reads cost ~10% of the base rate — about a 90% discount) for the system prompt and tool definitions you send on every call, and the Batch API (50% off input and output) for anything that isn't latency-sensitive, and the gap widens further.

What the August 31 cliff changes#

On September 1, 2026, Sonnet 5 output goes from $10 to $15 per million tokens and input from $2 to $3 — the introductory pricing simply expires. Re-run the reference agent at the new rate and the ~700 Sonnet drafts jump from ~$18 to ~$27 — call it $9 more per 1,000 tickets, a ~50% rise on the single biggest line item. The total moves from ~$28 to ~$37.

The subtler effect is on your escalation threshold. Today Sonnet is 2.5x cheaper than Opus on input and 2.5x on output, so escalating an ambiguous call to Opus is a real cost jump. After September 1, Sonnet is $3/$15 against Opus's $5/$25 — the ratio tightens, and the marginal cost of sending a borderline-hard call to Opus instead of Sonnet shrinks. That means two things: tighten your Sonnet usage now while it's cheap (cache aggressively, batch what you can), and re-check your Haiku→Sonnet and Sonnet→Opus thresholds on September 1, because the math that justified keeping something on Sonnet may now justify either dropping it to Haiku or escalating it to Opus.

The decision rule#

Route by the job: mechanical and high-volume → Haiku 4.5; reasoning, code, and tool loops → Sonnet 5; hard, rare, and expensive-to-get-wrong → Opus 5, sparingly. Default to Sonnet, push everything you can down to Haiku, and let Opus earn its place one escalation at a time. Do the Haiku offloading before August 31 so your Sonnet volume is already lean when the price climbs — then revisit the thresholds in September.

If you want the routing logic itself — a working three-tier router that classifies each call and dispatches it to the right model — that's the companion build: Build a three-tier Claude model router to cut your agent bill. For the week's news on the pricing move, see the Founders' Wire on Anthropic's price ladder, and if your agent is coding-heavy and you're weighing providers, the cross-provider coding cost comparison puts these numbers next to the alternatives.