Comparison

Claude Sonnet 5 vs Opus 5: which one to run

Both carry a 1M-token context. Opus 5 lists at 1.67× Sonnet 5’s official rate, but at 0.25× against 0.35× it bills 1.19× as much on both input and output. Here is when that buys you anything.

The difference is price, not capability class

Sonnet 5 and Opus 5 share the specifications people usually compare first. Both take a 1,000,000-token context, both emit up to 128,000 tokens, both support tool use and extended thinking, and both route through /v1/messages. Neither is a cut-down version of the other.

They part on price. Official rates are $3 per million input tokens for Sonnet and $5 for Opus; output is $15 against $25. The multipliers differ — 0.35× on Sonnet 5, 0.25× on Opus 5 — so you are billed $1.05 and $5.25 per million on Sonnet against $1.25 and $6.25 on Opus. Opus costs roughly 1.19× as much either way, not the 1.67× its official rates alone would suggest.

The practical consequence: on most coding work the two produce answers of the same shape, and the price difference is the whole difference. Where that stops being true is the subject of the next section.

What the price difference buys

On work Sonnet 5 already handles, the two return answers of the same shape, and the 1.19× is spent on nothing observable. That covers most editing, test-writing, and multi-step tool calling.

The gap shows up on long reasoning chains, where an early misstep propagates through every later step, and on work where a wrong answer costs more than the tokens do. Neither is a property of the task description — it is a property of your task, which is why the useful move is to run the same prompt through both ids and compare.

Both answer on the same endpoint with the same request shape, so switching is a one-field change and comparing costs you one extra call.

Both models answer on the same endpoint; only the model id changes.
curl https://token-share.app/v1/messages \
  -H "x-api-key: $TOKEN_SHARE_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Explain this stack trace."}]
  }'

What the 1M context does not do

A million tokens is room, not recall. Both models cost the same per token whether the window is nearly empty or nearly full, so pasting an entire repository into every request multiplies the bill without improving the answer.

Prompt caching is the lever that matters when a long prefix genuinely repeats — a system prompt, a file the conversation keeps returning to. Cached reads are billed separately from fresh input, so a stable prefix stops being charged at full rate on every turn.

Common questions

How do I tell whether Opus 5 is worth 1.19× on my work?
Run the same prompt through both ids and compare the outputs. On tasks Sonnet 5 already handles the answers tend to be comparable, and the extra spend buys nothing you can see. Where the difference tends to show is long reasoning chains and work where a wrong answer is expensive — but that depends on your task, not on a general ranking.
Do both models cost the same fraction of the official rate?
No. Sonnet 5 is billed at 0.35× the official per-token rates and Opus 5 at 0.25×, which is why Opus costs 1.19× as much rather than the 1.67× its official rates suggest. Multipliers are set per model: the catalog runs from 0.1× to 0.5×, and each model page shows its own.
Are these the same models Anthropic ships?
Yes. Requests are forwarded to Anthropic upstream; the model ids are theirs and the responses are theirs. What differs is billing and the account pool the request is served from.