Every platform that serves Claude — Bedrock, Vertex AI, Foundry, Claude Platform on AWS, and the first-party API — offers the same three-model lineup. The names are tiers, not versions: Opus is the most capable, Haiku is the fastest and cheapest, Sonnet sits between. Because all three speak the same API, switching tiers is a one-line change, which makes this one of the rare architecture decisions you can revise in an afternoon.
The lineup at a glance
| Model | Input / 1M tokens | Output / 1M tokens | Character |
|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | Most capable general model; the default recommendation |
| Claude Sonnet 5 | $3.00 ($2.00 intro through Aug 31, 2026) | $15.00 ($10.00 intro) | Balanced workhorse |
| Claude Haiku 4.5 | $1.00 | $5.00 | Fast and cheap; 200K-token context vs 1M for the others |
Prices are list prices per million tokens, and cloud marketplace list prices match the first-party API. Tokens are the billing unit — roughly word-fragments of text — and input (what you send) is always cheaper than output (what Claude writes). One structural difference hides in the last column: Haiku 4.5 has a 200K-token context window, while Opus 4.8 and Sonnet 5 take 1M tokens. If your workload involves very large documents in a single request, that alone can rule Haiku out.
Matching tiers to real workloads
Opus 4.8 is for work where quality failures are expensive: complex analysis and reasoning, legal or financial document review, multi-step agent workflows, code generation your engineers will actually merge. It is the default recommendation for a reason — when you don't yet know how hard your problem is, Opus tells you what "as good as it gets" looks like.
Sonnet 5 is the balanced choice for high-volume production work that still needs real intelligence: customer-facing assistants, document summarization, content drafting, general internal tooling. Its intro pricing through August 31, 2026 makes it especially cheap to evaluate right now.
Haiku 4.5 is for tasks that are high-volume, latency-sensitive, or simple enough that a bigger model is wasted: classification and routing, tagging, extraction from short documents, autocomplete-style features, first-pass triage of tickets and emails. At $1 per million input tokens, entire categories of automation become economically boring — which is a compliment.
Think in cost per task, not price per token
Per-million-token prices look abstract until you convert them into your actual unit of work. A short customer email and a one-paragraph reply consume a few hundred tokens each way — a small fraction of a cent even on Opus. A pipeline pushing millions of documents through daily is a different story, and the 5x input price gap between Opus and Haiku compounds fast at that scale. The right question is never "which model is cheapest?" but "what does one unit of this business process cost on each tier, and where does the quality difference stop paying for itself?"
Also remember the tiers are not mutually exclusive. Mature deployments routinely mix them — Haiku classifies and routes incoming work, Sonnet handles the standard cases, Opus takes the escalations. Because the API is identical, the routing logic is just a string swap in your code.
Two details that surprise people later
First, the choice is per-request, not per-contract. Nothing about buying through Bedrock, Vertex AI, Foundry, or Claude Platform on AWS commits you to a tier; every request names its own model, so a single application can use all three in the same afternoon. Budget approvals framed as "we're buying Opus" misunderstand the product — you are buying tokens, and the tier is a dial your engineers turn per task.
Second, intro pricing has a date on it. Sonnet 5's discounted rate runs through August 31, 2026, after which the standard $3.00/$15.00 list applies. If your cost model was built during the intro window, recheck it against standard pricing before signing off on a launch budget — a forecast that only works at the promotional rate is not a forecast.
Where to go next
Learn the billing unit itself in Tokens 101, then see cutting costs with model tiering for the mixed-tier pattern in practice. The model section of the main guide keeps the current lineup summary.