The price rise that won't happen, and the one already on the calendar

Introductory pricing is a promise with an expiry date attached, and the date is the part that ends up in your budget spreadsheet. In the space of four days this month, the two largest model vendors did opposite things with theirs.

The date that got cancelled

Anthropic's platform release notes for 2026-08-10 carry a single line: "The introductory pricing for Claude Sonnet 5 ($2 / $10 per MTok) is now the standard price: the previously scheduled increase to $3 / $15 per MTok on September 1, 2026 will not occur."

The pricing page repeats it as a standing note against the model table: "The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur."

That is a 33% reduction against the price that was on the calendar for next Tuesday, and it is the whole of the change — no new model, no tier, no commitment required. The table now lists Sonnet 5 at $2 input and $10 output per MTok, against Sonnet 4.6 and 4.5 at $3 and $15. Batch rates follow at half: $1 and $5.

For anyone who built a cost model in the last few months that assumed a September step up, the correction is straightforward and in your favor. It is also worth noting what did not change: the note governs Sonnet 5 only. Opus 5 remains $5/$25, Fable 5 and Mythos 5 remain $10/$50, Haiku 4.5 remains $1/$5.

Per token is not per task

Here is the wrinkle that a price table cannot show you, and that Anthropic's own pricing page puts three notes below the numbers:

Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer that contributes to their improved performance on a wide range of tasks. This tokenizer produces approximately 30% more tokens for the same text. The exact increase depends on the content and workload shape. Claude Sonnet 4.6 and earlier models use the previous tokenizer.

Claude Platform pricing documentation

Sonnet 5 is on the newer tokenizer; Sonnet 4.6 is on the previous one. So the cross-generation comparison is not $2 against $3. Running the documented figures: the same text costs 2 × 1.30 = 2.60 per notional MTok-of-old-tokens on Sonnet 5, against 3 on Sonnet 4.6. That is about 13% cheaper for the same work, not 33%. Output behaves the same way: 10 × 1.30 = 13 against 15, again roughly 13%.

Two caveats on that arithmetic, both from the source. The 30% is explicitly approximate and "depends on the content and workload shape" — code, prose, and non-English text will not move together. And it is a documented consequence of a capability change, not a hidden fee: the same note says the newer tokenizer "contributes to their improved performance."

The practical upshot is that per-token list prices are only comparable within a tokenizer generation. Comparing them across one — or across vendors, who do not share tokenizers at all — measures the wrong thing. If you want a real number, meter your own workload. On Anthropic's API the response usage object gives it to you per request.

Google shipped a fresh cliff

Four days after Anthropic removed an expiry date, Google added one. The Gemini API release notes for 2026-08-13 announce "Gemini 3.7 Flash generally available (GA)" as gemini-3.7-flash, described as the "most intelligent workhorse model yet for coding and agents" with "Substantial improvements across software engineering, web development, and agentic workflows."

The pricing page attaches introductory rates with a hard date. Input: "$0.75 through December 31, 2026. $1.50 starting January 1, 2027." Output: "$3.75 through December 31, 2026. $7.50 starting January 1, 2027."

That is a scheduled doubling, both sides, on New Year's Day. For comparison, the page lists Gemini 3.5 Flash at $1.50 input and $9.00 output. So the January reset puts 3.7 Flash at exactly 3.5 Flash's input price and below its output price — which reads less like a price rise than like a launch discount unwinding to where the line was already priced.

Anyone piloting 3.7 Flash on the introductory rate right now should model both columns. A workload that is comfortably economic at $0.75/$3.75 and marginal at $1.50/$7.50 has a review date, and it is not far away.

What to plan around

Three things follow from this pair.

First, treat "introductory pricing" as an option the vendor holds, not a schedule you can plan on in either direction. Anthropic let its date lapse into permanence with two weeks' notice; Google booked a new one four months out. Both are documented, both are in the primary sources, and neither was predictable from the previous quarter.

Second, put the expiry dates you are exposed to in the same place you keep certificate expiries. 2027-01-01 for anything on gemini-3.7-flash is now a real operational date.

Third, when a model generation changes tokenizers, re-meter rather than re-reading the table. A 33% headline cut that is 13% in practice is still a cut — but if you sized capacity or a per-customer margin on the headline, you sized it wrong.

Primary sources: Claude Platform pricing, Claude Platform release notes, Gemini API release notes, Gemini API pricing, read 2026-08-25.