Claude Opus vs Sonnet Cost Per Agent Turn

Turn-level comparison from Anthropic list rates · transcript re-read included

Tool calls per turn

Tokens per tool record (input + output)

System prompt (tokens)

Output share of a tool record (%)

Turns on Opus

Turns on Sonnet

Opus input $/MTok

Opus output $/MTok

Sonnet input $/MTok

Sonnet output $/MTok

Claude Opus

First turn-

Last turn-

Average turn-

Tokens re-read (total)-

Run total-

Claude Sonnet

First turn-

Last turn-

Average turn-

Tokens re-read (total)-

Run total-

How the comparison works

Both models face the same transcript. On turn i the model re-reads the system prompt plus every tool record from earlier turns, then reads and writes the new records for this turn. The formula per turn is

turn_cost(i) = ( sys + (i-1) * T * tok ) * pin / 1e6 + T * tok * (1 - outFrac) * pin / 1e6 + T * tok * outFrac * pout / 1e6

The run total sums this over the turn count you give each model. Prices are editable so you can check batch rates or future list changes; the defaults are the published Opus 5.5 and Sonnet 5 list rates.

Published list rates used by default

Model	Input $/MTok	Output $/MTok
Claude Opus 5.5	$4	$20
Claude Sonnet 5	$2	$10
Claude Haiku 4.5	$1	$5

Source: Anthropic pricing page, docs.claude.com, verified 2026-09-28. Prompt caching and batch multipliers are not applied on this page. See the prompt caching savings calculator for cache economics and the agent loop cost calculator for caching and retry effects inside one run.

Reading the break-even turn ratio

The verdict line reports the turn ratio at which Sonnet's total catches up to Opus. If Opus reliably finishes in fewer turns than that ratio, Opus is cheaper for the whole run. If not, Sonnet wins on cost and the difference goes to quality judgement, not arithmetic.

Same turn count on both models: Opus costs about 1.67x Sonnet at default rates (5/3 on input and on output), so Sonnet wins on cost alone.
Opus at half the turns usually still costs less than Sonnet at default rates. It approaches parity as the system prompt grows, because the big first-turn re-read lands on whichever model runs the first turn. Run your own numbers above.
Per-turn cost grows linearly for both models because the transcript re-read grows every turn. The last turn is always the most expensive one.

Frequently asked

Why does the last turn cost more than the first?

The transcript is longer. Every earlier tool record is re-read as input on each later turn, so per-turn cost rises linearly with turn number.

Does this include caching?

No. This page compares raw list-rate costs with the full transcript re-read priced at the input rate. Caching discounts the re-read portion and is modeled separately in the agent loop cost calculator.

Are the prices in the list exact?

They are the Anthropic published list rates as of 2026-09-28, and every field is editable, so you can substitute negotiated or batch rates.

Agent Loop Cost Calculator Prompt Caching Savings Agent Retry Cost Tool Use Cost Estimator All Tools