For a production workload in 2026, the price per token will shape your unit economics well before latency or answer quality does. Below: what drives the cost of Sonnet 4.6, Opus 4.8 and gpt-5.6-luna, where prompt caching genuinely cuts the bill, and how codexkey rates map to Anthropic's direct ones.
Three tiers, three cost shapes
The codexkey lineup splits into three tiers by how the bill is built:
- gpt-5.6-luna: the cheapest tier. Built for frequent, short calls such as classification, routing, extraction and light chat. Make it the default for any step of an agent pipeline that does not need long reasoning.
- Sonnet 4.6: the workhorse. Strong reasoning, a large context window, and a price above gpt-5.6-luna. Most production assistants and code agents end up here, because the quality-per-dollar curve is hard to beat.
- Opus 4.8: the frontier tier. Use it for hard reasoning, planning, and tasks where one Opus call replaces five Sonnet retries. Output tokens cost more than they do on Sonnet.
We carry Haiku too: the catalog has claude-haiku-5.5 and claude-haiku-4.5, with prices on the models page. They run in the Premium channel and are not available in the Cheap Haiku channel.
Input tokens vs output tokens
A common mistake when modelling Claude costs is to treat input and output tokens as one line item. They are two. Output tokens cost roughly 5× input tokens at every tier. For an assistant that writes long answers (summaries, code, reports), output sometimes takes 80% of the bill or more. For a RAG chat that returns short answers from a large context, input is the expensive side instead.
Two practical consequences:
- Cap
max_tokenshard. A 4 000 token ceiling on a chat that really needs 800 turns into a permanent 5× tax on the long tail of wordy answers. - Stream and bail out early. If the consumer can act on the first paragraph, cut the stream instead of waiting for the model to finish.
Prompt caching: where the savings are real
Prompt caching gives a large discount on cached input tokens (system prompts, long documents, tool schemas) on repeat calls. An agent loop that re-sends the same system prompt and tool list at every step can cut input cost by 80-90% at steady state. Turn it on.
How codexkey prices Claude
codexkey is a pay-per-token gateway to Claude. We match Anthropic's direct per-model rates, so you can change the base URL without re-modelling your costs. You pay only for the tokens you use: no monthly minimums, no seat fees, no commitment. Top up through SBP (the Russian Faster Payments System), crypto or card, and send requests to the same Anthropic-compatible or OpenAI-compatible endpoints your SDK already expects.
The live rate sheet is served at /v1/models/info, and the pricing page is linked
from the dashboard.
Bottom line
gpt-5.6-luna for cheap, Sonnet as the default, Opus for hard. Cap max_tokens
and cache your system prompts. After that, stop optimising and ship.