Sonnet 5 Stays at $2. Your Bill Might Still Go Up.

TL;DR

Anthropic has canceled its own price increase. Claude Sonnet 5 launched on June 30 at an introductory $2 per million input tokens and $10 per million output tokens, with standard pricing of $3/$15 scheduled to kick in on September 1. That 50 percent jump is now dead. The intro rate is permanent, per Anthropic's own announcement and an August 10 edit to the launch post. Before you refactor your cost model, note that Sonnet 5 uses the newer tokenizer that Anthropic's own docs say emits about 30 percent more tokens for the same text.

What actually changed

The original deal was a discount with an expiry date. Anthropic shipped Sonnet 5 in June, priced it aggressively for two months, and published a hard cutover: intro pricing through August 31, standard pricing from September 1. Every FinOps blog on the internet spent July telling you to audit your bill before the deadline.

The deadline is gone. Anthropic's post on X is unambiguous:

We're making Claude Sonnet 5's introductory pricing permanent. We launched Sonnet 5 in June at $2 per million input tokens and $10 per million output tokens through August 31, and that price will remain unchanged.

The launch announcement carries a matching edit dated August 10: the standard $3/$15 pricing "previously set to take effect September 1 no longer applies."

That puts Sonnet 5 permanently a third below claude-sonnet-4-6, which sits at $3/$15 and is not going anywhere. A newer, more agentic model that costs less per token than the one it replaces is not the usual direction of travel.

The line items nobody reads

The headline rate is the part people quote, but the freeze applies to the whole column, and if you run batches or lean on prompt caching that is where the money actually is.

• Batch API: stays at $1/$5 instead of moving to $1.50/$7.50.
• Cache hits and refreshes: stay at $0.20 per million instead of $0.30. On a cache-heavy agent loop, this is the line that decides your margin.
• Cache writes: the 5-minute write stays at $2.50 rather than $3.75, and the 1-hour write stays at $4 rather than $6.

All of those multipliers are fixed ratios off the base input price, so freezing the base froze everything downstream. Nothing here required a separate decision, which is presumably why nobody announced it.

Now the catch, and it is a real one

A per-token price is only half of a bill. The other half is how many tokens the job takes, and that number moved in the wrong direction.

Anthropic's pricing documentation carries a note that is easy to scroll past: Claude 4.7 and later models use a newer tokenizer, and it "produces approximately 30% more tokens for the same text." Sonnet 4.6 and earlier use the old one. Sonnet 5 is on the new one. So the same prompt, byte for byte, meters higher on Sonnet 5 than on its predecessor before a single word of output is generated.

This is shrinkflation run in reverse. The price per unit genuinely dropped, but the units got smaller, so the same job needs more of them. Anthropic is not hiding it, the note is right there in the docs, but it does mean the sticker comparison against Sonnet 4.6 flatters the new model.

Tokenizer inflation is the floor, not the ceiling. The Decoder has argued that Anthropic has a pattern of holding token rates flat while models consume more tokens per task, and its reporting puts Sonnet 5 at roughly 40 percent more output tokens per task than Sonnet 4.6 at maximum performance, running about three times as many agent loops. Those are third-party measurements rather than vendor figures, and they will vary wildly with your workload, but the direction is consistent with the tokenizer note.

The practical read: your per-token rate is now guaranteed, your per-task cost is not. Those are different promises, and only one of them was made.

Why Anthropic blinked

Nobody cancels a scheduled 50 percent increase because the spreadsheet demanded it. Look at where Sonnet 5 sits in the market and the logic writes itself.

Per OpenAI's pricing docs, GPT-5.6 Terra runs $2 per million input and $12 per million output. At $2/$10, Sonnet 5 matches Terra on input and undercuts it on output. At $3/$15, it would have been comfortably more expensive than the obvious competitor in the same tier, in the exact segment where switching costs are lowest and everyone has an abstraction layer with a model string in a config file.

The mid-tier workhorse model is the one that gets swapped on price, because it is the one running in a loop ten thousand times a day. Anthropic priced accordingly.

The docs have not caught up

As of publication, Anthropic's own pricing page still lists a "Claude Sonnet 5 starting September 1, 2026" row at $3/$15, complete with a footnote explaining that standard pricing "will take effect." If you are scraping that table into a cost model, you are currently scraping a price that no longer exists. The announcement landed on X and in the launch post first, which is a familiar order of operations.

What to actually do with this

• Cancel the September fire drill. If you had a migration or a budget review pinned to the August 31 deadline, it is moot. Nothing changes on September 1.
• Measure tokens, not rates. Pull real token counts per task from your own traces on Sonnet 5 and Sonnet 4.6. The rate card cannot tell you what a job costs when the tokenizer changed underneath it.
• Re-check the caching math. A cache hit at $0.20 per million is 10 percent of base input. If your agent replays a large system prompt, that ratio matters more than the headline rate.
• Do not hardcode $3/$15 from the docs. The stale row is still live.

A price cut you have to do arithmetic to feel is still a price cut. Just do the arithmetic on your own traces before you promise your CFO a third off.

Key Takeaways

• Anthropic made Claude Sonnet 5's introductory pricing permanent at $2 per million input tokens and $10 per million output tokens, canceling the $3/$15 standard pricing scheduled for September 1.
• The freeze extends to batch ($1/$5) and prompt caching (cache hits at $0.20 per million), because those rates are fixed multiples of base input.
• Sonnet 5 now sits permanently one third below Sonnet 4.6, which stays at $3/$15.
• Anthropic's docs state that Claude 4.7 and later models use a tokenizer producing about 30 percent more tokens for the same text, which partly offsets the lower rate against older Sonnets.
• At $2/$10, Sonnet 5 matches GPT-5.6 Terra on input and undercuts it on output; at $3/$15 it would have been the pricier option in its tier.
• Anthropic's pricing page still lists the canceled September 1 rates as of publication, so verify before wiring that table into a cost model.

Sources: Anthropic, Introducing Claude Sonnet 5 (August 10 edit), Claude on X, Anthropic pricing documentation, OpenAI API pricing, The Decoder