Claude Opus, Sonnet or Haiku: which tier to use and what each really costs
By LW Forge — maintainer of LLM Scout · Updated August 10, 2026
Most Claude cost questions aren't "Claude or GPT" — they're "which Claude". Once you've settled on Anthropic, you still have to pick between Opus 4.8, Sonnet 5 and Haiku 4.5, and that single choice moves your monthly bill by a factor of five on identical traffic. The useful part is that the ratio is almost perfectly clean, which makes the decision easier to reason about than it looks.
The price table
Rates per million tokens, as listed in July 2026 — the same figures behind the Claude cost calculator:
| Model | Input | Output | Cache read | Context |
|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | $0.50 | 200k |
| Claude Sonnet 5 | $3.00 | $15.00 | $0.30 | 200k |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | 200k |
Three things jump out. Output is billed at 5× input across all three tiers, so the output-heavy nature of your workload matters as much as the tier you pick. Cache reads land at roughly a tenth of the input rate everywhere, so the caching discount is tier-independent. And the context window is identical at 200k tokens — moving up a tier buys you capability, never more room, which is the single most common misconception about the lineup.
Sonnet 5 also carries a launch promotion of $2/$10 per million tokens through August 31, 2026. It's real money while it lasts, but don't build a budget on it: after that date the rate reverts to the $3/$15 shown above, and a pipeline sized on promo pricing gets a 50% increase overnight.
What that means for a real monthly bill
Take a support assistant handling 100,000 requests a month, each sending about 2,000 input tokens (system prompt plus retrieved context plus the user's message) and generating about 400 output tokens. That's 200M input and 40M output tokens a month:
| Model | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Opus 4.8 | $1,000 | $1,000 | $2,000 |
| Sonnet 5 | $600 | $600 | $1,200 |
| Haiku 4.5 | $200 | $200 | $400 |
Opus is exactly 5× Haiku and Sonnet is exactly 3× Haiku, because the tiers scale input and output by the same multiple. That's the practical takeaway: your tier choice is a constant applied to your entire bill, so you can estimate any scenario by pricing it once on Haiku and multiplying. Run your own token counts through the Claude cost calculator to see the version of this table that matches your traffic, and if you don't know your average request size yet, measure a representative prompt in the token calculator first.
Where each tier actually earns its rate
Haiku 4.5 is the right default for anything narrow and well-defined: classification, routing, tagging, extraction from a known schema, short summaries, moderation passes. These tasks have a verifiable right answer, which means you can measure quality directly instead of guessing — and at $1/$5 you can afford to run them at volume. The failure mode is subtle degradation on tasks that look simple but require judgment.
Sonnet 5 is the workhorse for production features where a user reads the output: chat assistants, drafting, code changes of moderate complexity, agents that chain a handful of tool calls. It handles ambiguity noticeably better than Haiku while staying cheap enough for real traffic. Most products that ship something customer-facing end up here.
Opus 4.8 earns its 5× premium on genuinely hard reasoning: multi-step refactors across a large codebase, dense analytical work where a wrong answer is expensive, long agentic runs where an early mistake compounds through every later step. The economics flip when the cost of a bad output exceeds the cost of the tokens — which is common for low-volume, high-stakes work and almost never true for high-volume, low-stakes work.
Two levers that beat tier-shopping
Before you pay for a bigger model, check whether you're paying full price for repeated text. Anthropic charges cache reads at about a tenth of the input rate, so a stable system prompt reused across requests is one of the largest discounts available at any tier — prompt caching covers when it works and when it quietly does nothing. The Batch API is the other easy win: roughly half price for work that doesn't need an immediate answer, which describes most classification and enrichment pipelines.
Mixed routing beats a single global choice more often than teams expect. Run the cheap tier first, escalate only the requests that fail a confidence check, and you pay the expensive rate on the fraction that needs it rather than on everything. The Sonnet vs Haiku breakdown works through that math on the most common pair.
Deciding without over-thinking it
Start one tier lower than instinct says, and let measured quality — not vibes — push you up. Because the ratio is constant, the question is always the same: is this tier's output more than 3× (or 5×) better for my specific task? For narrow tasks the honest answer is usually no; for open-ended reasoning it's often yes. Price your own scenario in the Claude cost calculator, and if you're still weighing providers rather than tiers, Claude Sonnet 5 vs GPT-5.6 and the cheapest LLM APIs of 2026 cover that comparison directly.