Claude Opus vs Sonnet vs Haiku — cost per finished task, not cost per token

The tier comparison everyone publishes is a table of token prices. The decision does not turn on token prices. It turns on cost per finished task — and a tier that costs five times as much per token but gets there in one attempt instead of three is not five times as expensive.

So this page gives you the arithmetic rather than a verdict, computed from live catalog rates with every input visible. The number it produces is not a price. It is an attempt count: how much worse a cheaper tier has to be before it stops being cheaper.

The rates

Rates are read from the catalog at render time and printed next to Anthropic's own list price, so both halves are checkable. Nothing on this page is a hand-typed number.

The arithmetic that actually decides it

# Cost per finished task, not cost per million tokens.
#
#   task_cost = step_cost x steps x attempts
#
# A step here is 25,000 input + 1,200 output tokens —
# one agentic round trip, sized the same way every cost page on
# this site sizes it.
#
#   Claude Haiku 4.5       $0.012 / step
#   Claude Sonnet 4.6      $0.037 / step
#   Claude Opus 5          $0.062 / step
#
# So the break-even is an ATTEMPT COUNT, not a price:
#
#   Claude Haiku 4.5 stops being cheaper than Claude Sonnet 4.6
#   once it needs 3.0x the attempts.
#
#   Claude Sonnet 4.6 stops being cheaper than Claude Opus 5
#   once it needs 1.7x the attempts.
#
# That is the whole decision. Everything else is judgement about
# how many attempts YOUR tasks take on each tier.

Read the two break-evens as questions about your own work, not as recommendations. If the cheaper tier gets your tasks right first time, it wins by a wide margin. If it needs a second and third pass — and worse, if a wrong answer costs you a review cycle rather than just a retry — it stops winning quickly.

Two things this model deliberately leaves out, because pretending otherwise would be dishonest. It does not price your time: a retry costs a step, and reading a wrong answer costs a person. And it does not include a measured quality difference between the tiers on your tasks, because that is workload-specific and we have not benchmarked yours. Both push in the same direction — toward the stronger tier — so treat the arithmetic as the floor of the case for Sonnet or Opus, not the whole of it.

Why the split beats the choice

Framing this as “which one model” is the mistake. Real workloads are mixed: the same coding session contains architecture decisions and commit messages, and paying Opus rates for a commit message is as wasteful as asking Haiku to design a schema.

# The setting most people never change, and the one that pays.
# Claude Code fires background sub-tasks on its own; this decides
# which model gets them.

export ANTHROPIC_MODEL=claude-sonnet-4-6
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5

# Same idea in any client with a per-role model field: strong model
# for the work, cheapest model for the housekeeping.

In Claude Code the background sub-tasks are frequent, automatic and invisible — which makes the Haiku mapping the highest-leverage line in the config and the one almost nobody sets. The same idea applies in any client with per-role model fields; CC Switch exposes it as an explicit Sonnet / Opus / Haiku mapping, and Continue as one entry per role.

Running the experiment

The attempt count is the only unknown, and it is cheap to measure because the tiers share a request format: run the same ten tasks on two tiers, count how many needed a second pass, and multiply. On Kunavo they also share one key and one base URL, so the experiment is an environment variable rather than a second account.

Put your own token counts into the Claude token cost calculator — including the cached fraction, which on agentic workloads is most of the input and shifts every number here. The full rate table is on the Claude API pricing page, and the subscription-versus-token question is on Claude Pro and Max limits.

FAQ