If you are choosing a Claude model for production workloads in 2026, the pricing differences between Sonnet 4.6, Opus 4.7, and Haiku will dominate your unit economics long before any latency or quality tradeoff does. This post breaks down what each model costs per 1M tokens, where prompt caching actually saves money, and how Claudexia's pricing maps to Anthropic's direct rates.
The three tiers, by cost shape
Anthropic's Claude family is intentionally split into three cost shapes:
- Haiku — the cheapest tier, optimised for high-throughput classification, routing, extraction, and lightweight chat. You should default to Haiku for any step in an agent pipeline that does not require long reasoning.
- Sonnet 4.6 — the workhorse. Strong reasoning, large context, and pricing roughly 5× Haiku. Most production assistants and code agents end up running on Sonnet because the quality-per-dollar curve is hard to beat.
- Opus 4.7 — the frontier tier. Reserved for hard reasoning, planning, and tasks where a single Opus call replaces five Sonnet retries. Expect output token prices around 5× Sonnet.
Input vs output tokens
A common mistake when modelling Claude costs is to treat input and output tokens as the same line item. They are not. Output tokens cost roughly 5× input tokens at every tier. For an assistant that emits long responses (summaries, code, reports), output dominates the bill — sometimes 80% or more. For a retrieval-augmented chat that returns short answers from a large context window, input dominates.
Two practical implications:
- Cap
max_tokensaggressively. A 4 000 token cap on a chat that really only needs 800 tokens is a permanent 5× tax on the long tail of verbose answers. - Stream and stop early. If your downstream consumer can act on the first paragraph, terminate the stream rather than letting the model ramble.
Prompt caching: where the real savings live
Prompt caching gives you large discounts on cached input tokens (system prompts, long documents, tool schemas) on subsequent calls. For agent loops that re-send the same system prompt and tool list on every step, caching can drop input cost by 80–90% at steady state. Use it.
How Claudexia prices Claude
Claudexia is a pay-per-token Claude gateway. We match Anthropic's direct per-model rates so you can switch base URL without re-modelling your costs. You pay only for tokens used — no monthly minimums, no seat fees, and no commitment. Top up via SBP, crypto, or card, and consume from the same Anthropic-compatible or OpenAI-compatible endpoint your SDK already expects.
The full live rate sheet is available at
/v1/models/info and the pricing page links from the dashboard.
Bottom line
Pick Haiku for cheap, Sonnet for default, Opus for hard. Cap your output tokens. Cache your system prompts. Then stop optimising and ship.