Pricing verified: 11 August 2026. Compiled from Anthropic’s official Claude Sonnet 5 announcement page and its published platform pricing documentation. We do not hold a paid Anthropic API account and have not benchmarked these models ourselves — every rate below is a published vendor number with its source named inline, and the arithmetic built on top of those rates is our own and labelled as such. Where Anthropic’s own pages disagree we say so rather than pick one silently.

Compiled and fact-checked by the AI Tools Worth editorial team. Corrections: contact page.

The 1 September price rise has been cancelled

Claude Sonnet 5 launched on 30 June 2026 at $2 per million input tokens and $10 per million output tokens, described at the time as introductory pricing that would run through 31 August. The plan was a step up to $3 / $15 on 1 September — a flat 50% increase on both halves of the bill.

That step up is no longer happening. Anthropic’s announcement page now carries an update dated 10 August 2026 stating that “Sonnet 5’s introductory pricing of $2 per million input tokens and $10 per million output tokens is now permanent.” The same update says the standard pricing of $3 input / $15 output originally scheduled to begin 1 September “no longer applies.”

If you built a Q4 budget around a September cost increase on Sonnet 5, that line item can come out. Nothing else about the model’s billing changed — this is a price that was scheduled to rise and now will not.

What Claude Sonnet 5 costs now, in full

The permanent rates, and the rates that were cancelled, side by side. All figures per million tokens (MTok), as published by Anthropic.

Rate	Permanent (from 10 Aug 2026)	Cancelled 1 Sept rate	Increase avoided
Base input	$2	$3	50%
Output	$10	$15	50%
5-minute cache write	$2.50	$3.75	50%
1-hour cache write	$4	$6	50%
Cache hits & refreshes	$0.20	$0.30	50%
Batch API input	$1	$1.50	50%
Batch API output	$5	$7.50	50%

The increase was uniform: every derived rate was set to rise by exactly half, because the cache and batch rates are fixed multipliers of the base input price. Anthropic publishes those multipliers directly — a 5-minute cache write is 1.25× base input, a 1-hour cache write 2×, a cache read 0.1×, and the Batch API applies a 50% discount to both input and output.

One limit on the above. Anthropic’s update names only the $2 / $10 base rates as permanent. The cache and batch rows in the table follow from the published multipliers rather than from a separate statement about each one, and that inference is ours.

What the cancelled increase would have cost you

The table below is our arithmetic on Anthropic’s published rates, not a vendor figure. Because the increase was a uniform 50%, the avoided cost is simply half of what you spend today on Sonnet 5, whatever your volume.

Monthly volume	At $2 / $10 (permanent)	At $3 / $15 (cancelled)	Avoided per month	Avoided per year
10M in / 2M out	$40	$60	$20	$240
50M in / 10M out	$200	$300	$100	$1,200
100M in / 20M out	$400	$600	$200	$2,400
500M in / 100M out	$2,000	$3,000	$1,000	$12,000
1B in / 200M out	$4,000	$6,000	$2,000	$24,000

These assume the base rates with no prompt caching and no Batch API discount. Both of those cut the absolute bill substantially, and both would have risen by the same 50%, so the proportional saving is unchanged.

Anthropic’s own pricing docs still show the increase

This is the part worth checking before you cite a figure at anyone. As of our check on 11 August 2026 — a full day after the announcement page was updated — Anthropic’s platform pricing documentation had not been updated to match. It still carried a dedicated table row reading “Claude Sonnet 5 starting September 1, 2026” at $3 base input and $15 output, immediately below the row labelled “Claude Sonnet 5 through August 31, 2026” at $2 / $10.

The explanatory note beneath that table still read, in full: “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.”

The Batch API table on the same page carried the same pair of rows — $1 / $5 through 31 August, $1.50 / $7.50 from 1 September. We re-fetched the page with a cache-busting query string to rule out a stale copy being served to us, and both readings were identical.

We are not treating this as a contradiction of the announcement. The documentation page itself defers upward, telling readers: “For the most current pricing information, visit claude.com/pricing.” The reasonable reading is a documentation page that has not yet caught up with a two-day-old pricing decision. The practical consequence is the point: a developer who lands on the pricing docs today will still be told to plan for a 1 September increase that has been called off. If a cost model at your company was built from that table this week, it is wrong in your favour.

Sonnet 5 against Sonnet 4.6: the headline gap is not the real gap

With $2 / $10 now permanent, Sonnet 5 is listed below its own predecessor. Claude Sonnet 4.6 stays at $3 / $15 — on the sticker, the newer model is a third cheaper than the older one on both input and output.

That comparison does not survive contact with Anthropic’s tokenizer note, which is on the same documentation page and is easy to miss: “Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer… 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.”

Anthropic does not name Sonnet 5 in that sentence. It names a boundary — 4.7 and later on the new tokenizer, Sonnet 4.6 and earlier on the old one — and Sonnet 5 sits above that boundary. Reading Sonnet 5 as using the newer tokenizer is our inference from that wording, not a direct statement. On that reading, a price per token is not comparable across the two models, because the same document produces a different number of tokens on each.

The arithmetic below is ours. It takes a fixed body of text, prices it on Sonnet 4.6 at $3 / $15, and prices the same text on Sonnet 5 at $2 / $10 after inflating the token count.

Token inflation on Sonnet 5	Effective input cost vs Sonnet 4.6	Effective output cost vs Sonnet 4.6	Verdict for the same text
0% (sticker comparison)	$2.00 vs $3	$10.00 vs $15	33% cheaper
~30% (Anthropic’s stated figure)	$2.60 vs $3	$13.00 vs $15	~13% cheaper
50%	$3.00 vs $3	$15.00 vs $15	Break-even
Above 50%	Over $3	Over $15	More expensive

Two things fall out of that. The real saving against Sonnet 4.6 at Anthropic’s own stated ~30% figure is around 13%, not 33%. And the break-even point is a 50% token inflation — comfortably above the stated ~30%, so Sonnet 5 should still be the cheaper of the two for the same work, but with far less headroom than the price list implies. Anthropic explicitly says the increase “depends on the content and workload shape”, so the only way to know your own number is to count tokens on your own text.

Timeline

Date	What happened
30 June 2026	Claude Sonnet 5 announced at $2 / $10, introductory through 31 August
10 August 2026	Announcement page updated: introductory pricing made permanent; the 1 September rates “no longer apply”
11 August 2026	Our check: platform pricing documentation still lists the 1 September increase in both the standard and Batch API tables

Frequently asked questions

Is Claude Sonnet 5 going up in price on 1 September 2026?

No. The increase was cancelled in an update dated 10 August 2026 on Anthropic’s announcement page, which states the $2 / $10 introductory pricing “is now permanent” and that the $3 / $15 standard pricing scheduled for 1 September “no longer applies.” Anthropic’s separate pricing documentation had not been updated to reflect this when we checked on 11 August.

What does Claude Sonnet 5 cost per million tokens?

$2 per million input tokens and $10 per million output tokens, now permanently. Prompt caching is $2.50 per million for a 5-minute cache write, $4 for a 1-hour write, and $0.20 for cache hits and refreshes. The Batch API halves both sides, to $1 input and $5 output.

How much would the cancelled increase have cost?

Exactly 50% more on every rate. On a workload of 100 million input and 20 million output tokens a month, that is $600 instead of $400 — $200 a month, or $2,400 a year, that is no longer being spent (our arithmetic on the published rates).

Is Sonnet 5 cheaper than Sonnet 4.6?

On the price list, yes — $2 / $10 against $3 / $15, a third less. For the same text the gap is narrower, because Anthropic states its newer tokenizer produces roughly 30% more tokens for identical input. At that figure the effective saving is about 13% (our arithmetic), and the two models would reach parity at 50% token inflation.

Why do Anthropic’s pricing pages disagree?

We do not know, and we are not asserting that either page is wrong on purpose. The announcement page was updated on 10 August; the documentation page still showed the old schedule on 11 August. The documentation page itself points readers to claude.com/pricing as the current source, which is consistent with a docs page lagging a recent decision.

Which Claude model is cheapest on the API?

Of the current models on the published table, Claude Haiku 4.5 at $1 input / $5 output per million tokens. Sonnet 5 sits above it at $2 / $10, and Claude Opus 5 at $5 / $25. We have not evaluated whether the cheaper model is adequate for any particular workload — this is a price ranking only.

What we could not verify

Gaps we could not close, listed here rather than papered over.

• Whether the cache and batch rates were formally made permanent alongside the base rates — the update names only $2 / $10. We derived the rest from the published multipliers.
• Whether Sonnet 5 uses the newer tokenizer — Anthropic’s note defines the boundary as “4.7 and later” and excludes “Sonnet 4.6 and earlier”, but never names Sonnet 5. Our reading follows the boundary.
• Your actual token inflation — Anthropic gives “approximately 30%” and says it varies by content and workload shape, with no range or ceiling.
• When the pricing documentation will be corrected — no published statement, and it still showed the 1 September rows when we checked.
• The API rate table on claude.com/pricing — the per-token API figures did not render for us on that page, so we could not use it as a third cross-check.
• Whether third-party platform pricing follows — Amazon Bedrock and Google Cloud set their own Claude rates, and we did not check whether either reflects the cancellation.
• Anthropic’s reasoning — no stated rationale for making the introductory rate permanent.

Sources and method

Compiled 11 August 2026 from Anthropic’s official “Introducing Claude Sonnet 5” announcement page (published 30 June 2026, carrying an update dated 10 August 2026) and Anthropic’s platform pricing documentation, read the same day. The documentation page was fetched twice, the second time with a cache-busting query string, to confirm the 1 September rows were genuinely still live and not a stale copy. Vendor rates and notes are quoted directly; all cost projections, percentage comparisons and the tokenizer break-even calculation are our own arithmetic on those published rates and are labelled where they appear. A post from Anthropic’s official social account was reported in search results as carrying the same announcement; we could not open it directly and have not relied on it. Prices change — verify against the vendor before you commit.