Don't Order Sonnet 5 Off a Cancelled Calendar

Late August is a bad week to trust a countdown.

On 10 August 2026, Anthropic said Claude Sonnet 5’s launch rate — $2 per million input tokens and $10 per million output tokens — would become the standard price. The planned move to $3/$15 on 1 September will not occur. Anthropic’s current Platform pricing page still shows the old dates, but only inside a note that explicitly cancels that schedule.

That distinction matters. This is a correction to a future plan, not a new price cut on September 1. Nothing is scheduled to get cheaper that day, and the cancelled higher rate must not be treated as a current fact.

A stale sheet could still make a team front-load Batch jobs or forecast September at $3/$15. If that calendar is in your model, close it. The rate that matters is Anthropic’s live list when you spend, read together with the dated cancellation — not a third-party table that copied the June launch window.

What the official record says

Anthropic’s Platform pricing table lists Claude Sonnet 5 at $2 / MTok input and $10 / MTok output. Five-minute cache writes are $2.50, one-hour cache writes are $4, and cache hits and refreshes are $0.20 per million tokens. The Batch table lists $1 input / $5 output.

Immediately below the model table, Anthropic preserves the history and resolves it: $2/$10 was announced as introductory pricing through 31 August, it is now the standard price, and the previously scheduled September 1 increase to $3/$15 will not occur. The dates remain visible as a cancellation record, not an active countdown.

The 30 June launch post has an Edit August 10, 2026 changelog. It says the $2/$10 rate is now permanent and the $3/$15 rate previously due on September 1 no longer applies. Anthropic’s August 10 social post gives the same update. “Permanent” here means the announced step-up was withdrawn; it is not a contractual promise that Anthropic can never change list prices again.

The main Claude pricing page now has an API table too, and its Sonnet 5 row agrees at $2/$10. The Platform docs remain useful for the complete cache and Batch breakdown. The current What’s new in Claude Sonnet 5 page also shows $2/$10 without the old step-up. These are now consistent first-party signals.

Open the official pages before committing spend. Anthropic can change list prices. The narrower dated claim is that, when checked on 28 August 2026, Sonnet 5 was $2/$10, Batch was $1/$5, and the former September 1 move to $3/$15 was explicitly cancelled.

Why an old countdown can still appear

Third-party rate cards often preserve a launch footnote after the first-party status changes.

TokenRate’s Sonnet 5 page, marked last verified 27 August 2026, still said $2/$10 applied through August 31 and then $3/$15, while its limitations section said the launch promotion expired on August 31. Its live dollar cells showed $2/$10; the attached calendar was stale.

BenchLM’s Claude API pricing hub, marked last synced 27 August 2026, also described $2/$10 through August 31 followed by $3/$15. Its Sonnet 5 row labeled the effective window “Through Aug 31, 2026.” The values matched the current list, but the expiry condition did not reflect Anthropic’s August 10 cancellation.

Neither page invented the original promotion. Anthropic published that window in June, aggregators copied it, and Anthropic cancelled the step-up on August 10. A recent “verified” or “synced” stamp is not enough if the page still presents an overridden future event as active.

Treat a page that says the September step-up will happen as a snapshot of the June plan. A first-party page may still mention the same dates accurately when its operative words are standard price, will not occur, or no longer applies. Read the status, not merely the presence of a date.

How to tell a current list from a stale sheet

Use three checks in order.

Read the official model row. Open Anthropic’s pricing table and find Claude Sonnet 5. On 28 August, the list was $2/$10 and Batch was $1/$5. If a spreadsheet disagrees with those cells, do not average the two or choose the higher figure as a vague form of conservatism. Reconcile it against the first-party table.

Read the verb attached to the date. “Through August 31, then $3/$15” describes a live step-up. “The scheduled increase will not occur” cancels it. A page can show the same dates and dollars while expressing opposite billing states. The predicate decides whether the calendar is active.

Check the dated change record. The August 10 edit on the launch post says the introductory price became the standard and the higher rate no longer applies. Use the changelog to establish when the plan changed, then use the live table for the rate you budget.

Do not substitute a June news clip, a partner SKU on another cloud, or an aggregator’s verification stamp for the current first-party table. Partner prices can differ from first-party Anthropic; that is a different invoice, not a preview of September.

If two pages disagree, compare recency and operative status rather than assuming every date is a future commitment. A historical note that says “will not occur” overrides a copied countdown that says “then $3/$15.” Re-open the list the day you commit spend.

What the cancelled step-up means for money

A cancelled increase is not a discount to harvest before midnight. It is a non-event.

If you already queued Batch jobs to beat August 31, you do not need to drain the queue or add work because of this former deadline. Batch is 50% of the current list, or $1/$5, exactly as the pricing page shows. There is no first-party September 1 cliff to front-run. Run jobs in the ordinary operational window.

If a September forecast still uses $3/$15, rewrite the rate line. Input at $3 instead of $2 is 50% higher; output at $15 instead of $10 is also 50% higher. Batch at $1.50/$7.50 instead of $1/$5 is the same overstatement. A workload costing $40,000 a month at the current list becomes a false $60,000 forecast under the cancelled rate. The $20,000 difference is not savings captured by stocking up; it is a model error.

If you have not front-loaded work, do not do so for this date. Tokens are not warehouse inventory. Pulling work forward usually means running jobs before the workload is ready, for a rate that is not scheduled to change.

Keep tokenizer effects in a separate forecast line. Anthropic’s launch post says Sonnet 5’s updated tokenizer can produce roughly 1.0–1.35 times as many tokens for the same text depending on content; the current Platform pricing page summarizes the increase as approximately 30%. That can raise a real bill without changing the $2/$10 sticker. It is a unit-count issue, not a September rate increase.

The operating rule is narrow: do not treat a withdrawn calendar as live, and do not let a third-party countdown override a first-party cancellation statement.

The decision

• Verify the price: open Anthropic’s Platform pricing table, read the Sonnet 5 row, and confirm the note says the September increase will not occur.
• Handle queued Batch normally: current Batch is $1/$5, with no official September 1 step-up to beat. Do not add volume merely to front-run a cancelled date.
• Trust a third-party table only after reconciliation: it should match the first-party rate and the latest official status, without presenting the withdrawn $3/$15 schedule as a future event.