Claude Sonnet 5 Pricing Became Standard: What $2/$10 Means for Production Routing

What actually happened on September 1

When Sonnet 5 launched, Anthropic published $2 per million input tokens and $10 per million output tokens as introductory pricing valid through August 31, 2026, with $3/$15 as the standard rate thereafter. That framing propagated widely into cost models, internal planning documents, and comparison tables that quietly encoded a 50% input increase and a 50% output increase as a scheduled event.

The increase did not happen. Anthropic's current pricing page lists Sonnet 5 at $2/$10 with no introductory qualifier and no future step-up date. The practical effect is that a large amount of published cost analysis — including the July baseline of this site — carried a price that never came into existence.

This is a small fact with a disproportionate footprint, because pricing pages are read once and cached in spreadsheets for quarters. It is worth re-checking any routing threshold, per-task budget cap, or build-versus-buy model that was calibrated between June and August 2026.

Permanent prices change routing, temporary ones do not

A discount with an expiry date is not a design input. Teams rationally avoid architecting around it, because the cheap tier becomes expensive on a known date and the migration cost lands later. A standing price is different: it can carry a default route.

At $2/$10, Sonnet 5 sits at 40% of Opus 5's input price and 40% of its output price, and at 20% of Fable 5.1's input price. It shares the 1M-token context window and 128K maximum output with both larger models. That combination — frontier-class capacity at a fraction of frontier price — is what makes a volume tier viable rather than merely cheap.

The routing question is therefore no longer "can we tolerate Sonnet 5 until the price rises," but "which classes of request genuinely require escalation." That is a much healthier question, and it is answerable with an eval rather than a calendar.

Where Sonnet 5 is the wrong answer

Anthropic's guidance places Sonnet 5 as the speed and intelligence balance and directs demanding reasoning and long-horizon agentic work to Fable 5.1, with Opus 5 as the general starting point. Its own launch-table evidence, discussed in the lineup report, shows large tier gaps on science and automation benchmarks and much smaller ones on some coding rows.

The honest read is that Sonnet 5 is a strong default for high-volume, well-specified, short-horizon work, and a poor default for tasks that require many hours of autonomous operation, deep scientific reasoning, or recovery from unexpected environment states. The failure mode is not a slightly worse answer; it is more retries, more human repair, and more tokens spent reaching the same place.

That is why cost per accepted result, not cost per token, is the only metric that settles this. A model at 40% of the price that needs 2.5x the attempts has saved nothing and added latency.

Rebuild the cost model, not just the price constant

Three other levers move the invoice more than the rate card does. Prompt caching on most Claude models prices cache reads at 10% of base input. The Batch API is 50% off for work that tolerates delay. And effort settings change token spend materially on models with adaptive thinking, which is why Anthropic argues that tuning effort is often a better lever than switching models.

Note that the cache-read discount is not uniform across the lineup. Fable 5.1 cache reads are priced at 2.5% of base input, or $0.25 per million tokens. For cache-heavy agent traffic this narrows the Fable-to-Sonnet gap in a way a rate-card comparison hides entirely, so the tier decision should be made on a workload trace rather than on the price column.

A defensible routing design measures each tier on the same task set, at a named effort setting, with caching configured the way production will configure it, and reports cost per accepted result alongside latency. Our cost comparator is a starting estimator for that work, not a substitute for it.

FAQ

Cite this page