TEARDOWN Nº 3 / CORRECTED

The Sonnet 5 price cliff that never happened

PUBLISHED 2026-08-13 · BY Llorenç Ballester — Chief Burn Officer

Claude Sonnet 5's introductory price was documented to expire on 2026-08-31, taking input from $2 to $3 per million tokens and output from $10 to $15. We published a teardown pricing that hike. Hours later we found Anthropic had cancelled it. This is the corrected version: what the increase would have cost, and the more useful story — how a number nobody touched went stale and got published anyway.

What actually happened

Sonnet 5 launched with introductory pricing of $2 input / $10 output per million tokens, announced as running through 2026-08-31, after which it would revert to a $3/$15 list price. That was real, documented, and is exactly what our board recorded. Anthropic has since made the introductory rate permanent and cancelled the increase. The price you pay today is the price you will pay in September.

What the cancelled increase would have cost

Worth keeping, because it sizes what the decision is worth to anyone running Sonnet 5 at volume. Each row uses that workload's own published assumptions from /workloads, with cache rates scaled by Anthropic's published multipliers.

Counterfactual figures — this increase never took effect and never will. Prices from the board, last confirmed 2026-10-14. Estimates at list prices, not invoices.

The cache rates, correctly this time

The first version of this page said Anthropic publishes no forward cache pricing, and presented our own extrapolation as the best available. That was wrong and worth correcting on its own. Anthropic documents the multipliers explicitly, relative to each model's base input price: 0.1x for a cache read, 1.25x for a 5-minute cache write, and 2x for a 1-hour cache write. Every Anthropic model on our board matches those multipliers exactly. The practical consequence, straight from their docs: caching pays for itself after one read on the 5-minute TTL, and after two on the 1-hour TTL.

How our own source went stale

This is the part we would rather write about than the price. Our pricing file carries a free-text note per model. Sonnet 5's said, in substance, “intro pricing, list is $3/$15 until 2026-08-31”. It was accurate when a human typed it in July. Nothing kept it accurate afterwards.

Our daily watcher cross-checks input and output prices against a public dataset every morning and stamps each model with the date it was last confirmed. It does that job well — Sonnet 5's numbers were correct the entire time. It has never read the note. So the prose sat there asserting a future that had been cancelled, next to numbers that were perfectly current.

The uncomfortable detail: we had a guard test for exactly this field. It asserted that the note still contained “$3”, “$15” and the date — so that if the terms changed, a human would be forced to update the constant instead of letting it drift. It passed green the whole time, because the note had not changed. What had changed was the world. A test that checks a file against itself can only detect edits, never obsolescence.

What replaced it is deliberately more modest, because the honest version of this lesson is that no automated check can verify an external fact — only the source can. The new guard refuses to let any model note reference a date that has already passed, which is the most common way hand-written provenance rots. Everything with a real expiry now has to be re-read at the source rather than trusted because it is written down. This page exists in its corrected form as the receipt.

If you budget on announced AI prices

Two things generalise beyond us. First, a scheduled price change is a plan, not a fact: this one was published, dated, and then withdrawn, and the withdrawal was quieter than the announcement. If a future price is load-bearing in your budget, re-read the source near the date rather than trusting your note from three months ago — which is precisely the mistake documented above. Second, if you keep any internal record of vendor pricing, ask what actually refreshes it. A number that updates automatically sitting next to prose that never does is a specific and very quiet trap.

Method and limits

Figures are computed at build time from the pricing board and the canonical workloads at /workloads. The cancellation and the cache multipliers are quoted from Anthropic's official pricing documentation, read on 2026-08-13. The counterfactual column describes a price that was never charged and is labelled as such. If Anthropic revisits the decision, this page is wrong again and we would rather you check the live board than take a dated article's word for it — which is, after all, the entire point of this correction. Full rules at /methodology.

Frequently asked questions