Cheaper agent models aren't cheaper until you benchmark the task

Claude Sonnet 5 cut sticker prices for agent workloads and changed the tokenizer. The only number that matters to a buyer is cost per completed task. Here is how to measure it before you standardize.

STICKER $/MTOK −33% TOKENS/TASK +UP TO ~35% PRICE THE TASK, NOT THE TOKEN COST / COMPLETED TASK = THE ONLY REAL NUMBER

On June 30, 2026, Anthropic released Claude Sonnet 5, a mid-size model tuned for agentic work (planning, driving browsers and terminals, running autonomously on tasks that previously needed larger, costlier models) at introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, rising to $3/$15 after (Anthropic announcement; TechCrunch, Jun 30, 2026). The same coverage carried a caveat buyers should not skim past: a new tokenizer can produce up to roughly 35% more tokens for the same text (TechCrunch, Jun 30, 2026). Those two facts together are the whole story. A price per token means nothing until you know how many tokens your task now costs.

Sticker price is the wrong unit

Model pricing is quoted per million tokens because that is what the vendor meters. It is not what you buy. You buy completed tasks: a qualified lead, a reconciled report, a triaged inbox, a merged pull request. Between the meter and the outcome sit variables the price sheet never shows: how many tokens the tokenizer cuts your documents into, how many steps the model needs, how often it retries, how often a human has to finish the job it started.

A tokenizer change moves the first variable directly. If the same contract, transcript, or codebase tokenizes 35% longer, a 33% sticker cut can wash out entirely on input-heavy work, and agent workloads are input-heavy by nature, because every step re-reads context. Whether the wash-out happens depends on your documents and your task shapes, which is precisely why nobody can answer it from the pricing page. Neither can we; the honest position is that the number is unknowable without your data in the harness.

Cost per completed task: the only benchmark worth running

Before standardizing any automation on a "cheaper" model, this one or its competitors', run the boring experiment:

Two additional wrinkles from the launch terms. The introductory pricing ends August 31, 2026, and the permanent rate is 50% higher. Benchmark at $3/$15, not $2/$10, or you are measuring a promotion, not a product. And because Sonnet 5 became the default on consumer plans the same day (MacRumors, Jun 30, 2026), your team's casual impressions of it will come from chat use, which tells you nothing about how it behaves on hour-long tool-using runs.

What this means for a mid-market buyer

The direction of travel is real and favorable: agent-capable models are getting cheaper faster than most buyers' evaluation practices are getting better. That gap is where money is lost. Firms with a benchmark harness, even a crude one, can switch models opportunistically every time the price-performance frontier moves, and it is moving every quarter now. Firms without one either stay loyal to an incumbent out of caution or chase sticker prices into cost overruns, and both are avoidable.

Building that harness is a two-week fixed-scope project, not a platform investment: task corpus, runner, token accounting, a pass/fail rubric a non-engineer can apply. It pays for itself the first time it stops a premature migration, or greenlights one with numbers instead of vibes. That is the work we would scope here, and it is the work we run on our own automation before we trust a vendor's arithmetic.