Blog · 2026-07-16 · Vynaris Team

Claude Sonnet 5 introductory pricing ends August 31, 2026: the 50% rise and who should re-route

Claude Sonnet 5's price rises 50% on Sep 1, 2026: input $2 to $3, output $10 to $15, batch and cache too. Per-task before/after, why the tokenizer makes it ~30% dearer than Sonnet 4.6 despite the same sticker, and who should re-route. Verified 2026-07-16.

Claude Sonnet 5's price rises 50% on September 1, 2026. Anthropic's pricing page states it plainly: "Introductory pricing of $2/$10 per million input/output tokens is in effect through August 31, 2026, after which the standard pricing of $3/$15 per million input/output tokens will take effect." This is not a discount ending on one line item. The entire Sonnet 5 price sheet, standard, batch, and cache, scales 1.5x on the same day. Prices verified 2026-07-16.

Frame it correctly before you plan around it: this is a 50% rise, not "50% off ending." The intro price is a 33% discount off the standard, and when it lapses the standard is 50% more than the intro. Same fact, but the direction matters when you are writing the budget.

TL;DR

• Every Sonnet 5 rate rises exactly 50% on Sep 1: input $2 to $3, output $10 to $15, batch $1/$5 to $1.50/$7.50, cache reads $0.20 to $0.30. An agent step (50k in / 15k out) goes from $250 to $375 per 1,000 tasks.
• The tokenizer twist: Sonnet 5 uses the newer tokenizer that emits ~30% more tokens for the same text than Sonnet 4.6, so per equal-text task Sonnet 5 crosses from ~13% cheaper than 4.6 to ~30% dearer, on an identical $3/$15 sticker.
• Re-route candidates on Sep 1: high-volume, quality-tolerant work (extraction, classification, routing) where Haiku 4.5 at $1/$5 or a cheaper model already suffices; keep Sonnet 5 where its quality earns the new price.

What actually changes

The whole sheet moves by the same multiplier. Here is the per-task bill before and after, prices verified 2026-07-16 from the Anthropic pricing docs, token counts held constant so you see the pure sticker change:

Workload                                     Intro /1,000  From Sep 1 /1,000  Rise
-------------------------------------------  ------------  -----------------  ----
Extraction (2k in / 200 out)                 $6.00         $9.00              +50%
Chat turn (3k in / 500 out)                  $11.00        $16.50             +50%
Cache-heavy step (45k cached + 5k / 2k out)  $39.00        $58.50             +50%
Agent step (50k in / 15k out)                $250.00       $375.00            +50%
Agent step, Batch API                        $125.00       $187.50            +50%

There is no shape that escapes it. Batch still halves the bill, prompt caching still cuts cache reads to a tenth of input, but both discounts now compute off a base that is 50% higher. If you priced a workload on Sonnet 5 this summer, multiply the per-token line by 1.5 and that is your September run rate.

The twist nobody put on the pricing page

Compare Sonnet 5's standard sticker to Sonnet 4.6: both are $3/$15. Identical. So a September migration from 4.6 to 5 looks free.

It is not, because they count tokens differently. Anthropic's own note: the newer tokenizer used by Sonnet 5 "produces approximately 30% more tokens for the same text. The exact increase depends on the content and workload shape." Sonnet 4.6 uses the previous tokenizer. So the same English prompt and the same English answer bill ~30% more tokens on Sonnet 5 than on 4.6, at the same per-token price.

Put the sticker rise and the token inflation together, on the same agent-step task, and Sonnet 5 crosses the Sonnet 4.6 line on Sep 1:

Version                                 Effective /1,000 tasks  vs Sonnet 4.6
--------------------------------------  ----------------------  -------------
Sonnet 4.6 (previous tokenizer)         $375                    baseline
Sonnet 5 intro (~30% more tokens)       $325                    13% cheaper
Sonnet 5 from Sep 1 (~30% more tokens)  $488                    30% dearer

The ~30% is content-dependent, so treat it as a band: at +20% more tokens Sonnet 5 lands 20% dearer than 4.6 on Sep 1; at +40%, 40% dearer. Either way the direction holds, and the sticker hides it. We worked the tokenizer math in full in Anthropic's tokenizer makes "same price" models ~30% more expensive; the Sep 1 rise stacks on top of it. Measure your own inflation with the count_tokens endpoint before you assume 30%, and price both versions on your own token counts rather than trusting the matching sticker.

Who should re-route on September 1

Not everything. The decision is per workload, not per account.

• Re-route away from Sonnet 5: high-volume, quality-tolerant jobs, extraction, classification, tagging, simple routing, where the answer quality gap to a cheaper model is small. Haiku 4.5 at $1/$5 is one-third of Sonnet 5's new $3/$15 and does not carry the same September step. On an extraction task the difference is $9 versus $3 per 1,000 tasks, and if quality holds, that is a clean 67% cut.
• Stay on Sonnet 5: work where its quality directly earns the 50% rise, complex reasoning, agentic coding, anything where a cheaper model's error rate costs more than the token savings. Paying more for the right answer is not waste.
• Re-examine Sonnet 4.6: for tasks already on 4.6, the September migration to Sonnet 5 is a ~30% per-task increase disguised as a lateral move. If 4.6 quality is adequate, "upgrading" is a price rise. Decide on measured quality, not version number.

The honest tradeoff: the intro window makes this the wrong month to lock a long contract on Sonnet 5's sticker, but it is also the right month to benchmark it while it is cheap. Run your evals now, at $2/$10, and know your Sep 1 number before your finance team asks.

The move

This is one instance of a pattern we keep finding: the pricing-page number is a floor, and the real per-request bill is set by things a sticker does not show, modifiers that stack, tokenizers that inflate, and dates that flip a rate overnight. The defense is the same each time. Meter the tokens you actually send, price each workload shape at the rate it will pay in September, and make the cheap-vs-quality call per request, which is a model-routing decision. The full July 2026 price list has the base rates every one of these adjustments sits on top of.

FAQ

Does Sonnet 5 get more expensive or cheaper on September 1?

More expensive. Input goes $2 to $3, output $10 to $15, a 50% rise across the whole sheet. The intro pricing was a 33% discount off standard; when it ends you pay 50% more than the intro rate.

Is Sonnet 5 still cheaper than Sonnet 4.6 after September 1?

No. Same $3/$15 sticker, but Sonnet 5's newer tokenizer emits ~30% more tokens for the same text, so per equal-text task it runs ~30% dearer than 4.6 on Sep 1. Today, at intro pricing, it is ~13% cheaper.

Do the batch and cache discounts still apply?

Yes, and at the same multipliers, but off the higher base. Batch is still 50% off ($1.50/$7.50 from Sep 1), cache reads still 0.1x of input ($0.30 from Sep 1). The discounts hold; the base they discount is 50% higher.

What should I do before August 31?

Benchmark Sonnet 5 while it is cheap, measure your real token inflation with count_tokens, compute your Sep 1 per-task cost, and identify the quality-tolerant workloads you can move to Haiku 4.5 or a cheaper model if the new rate does not pay for itself.

Vynaris is an OpenAI-compatible inference gateway that routes each request to the cheapest right-sized model and shows its receipts per request. One base URL swap. Get an API key at [vynaris.com](https://vynaris.com) and model your Sonnet 5 September exposure across workload shapes in the [calculator](https://vynaris.com/calculator). All prices verified 2026-07-16 against Anthropic's live pricing page; we will refresh this post at the September 1 changeover.