Claude Sonnet 5 Price Increase: I Priced 15 Days of Tokens

Anthropic's introductory pricing for Claude Sonnet 5 ends on August 31, 2026. From September 1 the standard card applies: $3 per million input tokens and $15 per million output, up from $2 and $10. The official pricing page already lists both rows, and every line moves by the same factor, including the cache rates that most coverage skips. A 50 percent increase is an abstraction until it lands on a workload you can measure, so I took 15 days of transcripts from the Mac mini that runs this site 24/7 and priced the same token stream at both cards.

The short version: $287.19 becomes $430.78, or $19.15 a day becomes $28.72. And the number the headlines lead with, the $2-to-$3 base input rate, accounts for 15 cents of the larger bill.

What changes on September 1

The docs note is one sentence: “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 full card matters more than the headline pair:

Rate	Through Aug 31	From Sept 1
Base input	$2 / MTok	$3 / MTok
5-minute cache write	$2.50 / MTok	$3.75 / MTok
1-hour cache write	$4 / MTok	$6 / MTok
Cache hit	$0.20 / MTok	$0.30 / MTok
Output	$10 / MTok	$15 / MTok
Batch input / output	$1 / $5	$1.50 / $7.50

One sentence from the launch announcement is worth keeping in view: “The introductory pricing is set so that the transition to Sonnet 5 is roughly cost-neutral.” Cost-neutral relative to Sonnet 4.6, and only at the introductory rate; that sentence has an expiry date.

Repricing 15 days of a real fleet

Claude Code keeps a JSONL transcript of every session under ~/.claude/projects/. I have measured this corpus before and stepped on both of its landmines, documented in my post on Claude Code token usage in JSONL: you must deduplicate by message.id because Claude Code writes one record per content block, and input_tokens is only the slice after the last cache breakpoint, not the input total. Same method, fresh pass, run this morning: 136 transcripts covering July 21 to August 4, deduplicating to 3,824 API responses.

• cache_read_input_tokens: 673,266,080, which is 95.9 percent of the input side
• cache_creation_input_tokens: 28,543,293, all of it at the 1-hour TTL that subscriptions get automatically
• input_tokens: 50,152, or 0.007 percent of the input side
• output_tokens: 3,826,323

Priced at both Sonnet 5 rate cards:

Component	Intro (through Aug 31)	Standard (Sept 1)
Cache reads	$134.65	$201.98
1-hour cache writes	$114.17	$171.26
Output	$38.26	$57.39
Base input	$0.10	$0.15
Total, 15 days	$287.19	$430.78

Two honesty notes before anyone quotes these numbers. First, this is a shadow price, not a bill: this fleet runs Opus 5 and Fable 5 on a flat-rate subscription, so what I am pricing is the question what would this exact token stream cost as Sonnet 5 on the API. The counts transfer cleanly because every model in the corpus already uses the newer tokenizer. Second, transcript-derived numbers are a floor, not a total: the transcript does not contain every request Claude Code makes, a limit I went through in the measurement post.

For a single unit of work the numbers get concrete. In the July 31 pass, 32 scheduled blog-publishing runs consumed 189.9M cache reads, 4.96M cache writes, and 1.69M output tokens. At Sonnet 5 rates, publishing one post on this rig prices out at $2.34 today and $3.50 from September.

The $3 headline is the least informative line

Every line scales by exactly 1.5, so the total moves by exactly 50 percent regardless of workload shape. What the composition decides is which line is worth optimizing. On this fleet, caching is 86.7 percent of the bill: cache reads alone are 46.9 percent and cache writes 39.8. Output is 13.3 percent. Base input, the line every headline quotes, is 0.03 percent.

The reason is structural to agent loops, not particular to mine. Each turn re-reads a context that each turn grows; this corpus reads about 176 cached input tokens for every output token it produces. So for agent workloads the rate to watch is not $2 going to $3. It is $0.20 going to $0.30 per million cache-read tokens, applied to the largest and fastest-growing number in the transcript.

Coming from Sonnet 4.6, it is more than 50 percent

The pricing page carries a second note that compounds with the increase: “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 announcement gives the range as “roughly 1.0–1.35× depending on the content type.”

Put the two notes together. Through August, moving from Sonnet 4.6 to Sonnet 5 was roughly cost-neutral by design: a lower rate offsetting a heavier token count. From September 1, Sonnet 5 carries the same $3/$15 list price as Sonnet 4.6 while metering up to 1.35 times the tokens for the same text. A job that cost $100 on Sonnet 4.6 lands somewhere between $100 and $135 on Sonnet 5, and either way it costs 50 percent more than it does this month.

The community did this arithmetic at launch. In the Hacker News thread on Sonnet 5 (1,266 points, 784 comments), one commenter wrote that Anthropic “introduced the new tokenizer to increase token generation by upto 33%”, and a heavily upvoted line of discussion held that cost per task “rises above opus at anything higher than medium effort.” I cannot verify the per-task claim from my transcripts, since this fleet does not run Sonnet 5. The tokenizer figure, though, is Anthropic's own documentation, not a community estimate.

What to check before September 1

If you bill through the API, four weeks of the intro card remain. The useful move is to price your own transcripts instead of reasoning from headlines. This is the whole script:

import json, glob, os

CARDS = {"intro": (0.20, 4.00, 2.00, 10.00),
         "standard": (0.30, 6.00, 3.00, 15.00)}  # read, 1h write, input, output

seen = {}
for f in glob.glob(os.path.expanduser("~/.claude/projects/*/*.jsonl")):
    for line in open(f):
        try:
            rec = json.loads(line)
        except ValueError:
            continue
        m = rec.get("message") or {}
        if rec.get("type") == "assistant" and m.get("id") and m.get("usage"):
            seen[m["id"]] = m["usage"]   # dedup: one record per content block

rd  = sum(u.get("cache_read_input_tokens") or 0 for u in seen.values())
wr  = sum(u.get("cache_creation_input_tokens") or 0 for u in seen.values())
raw = sum(u.get("input_tokens") or 0 for u in seen.values())
out = sum(u.get("output_tokens") or 0 for u in seen.values())

for name, (r1, r2, r3, r4) in CARDS.items():
    print(f"{name}: ${(rd*r1 + wr*r2 + raw*r3 + out*r4) / 1e6:,.2f}")

It prices all cache writes at the 1-hour rate, which is exact for subscription transcripts. If you call the API with 5-minute TTLs, split cache_creation.ephemeral_5m_input_tokens and ephemeral_1h_input_tokens and rate them separately. Beyond measuring, the levers I can see from here:

• Batch API keeps the 50 percent discount shape. $1.50/$7.50 from September. If your runs are not latency-sensitive, that single change returns you to almost exactly the intro card.
• Cache TTL is a real choice on the API. The 5-minute write costs $3.75 against $6 for 1-hour, but a shorter TTL that misses forces re-writes of the whole prefix, so whether it wins depends on the gap between your calls. Subscriptions get the 1-hour TTL automatically, with no choice.
• Sonnet 4.6 stays at $3/$15 on the previous tokenizer. From September it is the cheaper Sonnet per unit of text, at whatever capability cost your workload actually notices. That comparison is worth running before the default upgrade path decides for you.
• Subscribers see none of this on an invoice. Pricing reaches a subscription as limits, and limits fail differently: last week the Claude Code weekly limit killed 28 unattended runs on this rig with no warning an unattended process could read.

FAQ

When does Claude Sonnet 5 pricing change?

Introductory pricing of $2/$10 per million input/output tokens runs through August 31, 2026. Standard pricing of $3/$15 takes effect September 1, 2026. Cache and batch rates rise by the same 50 percent at the same moment.

Does the price increase affect Claude Code subscribers?

Not on an invoice. Pro and Max subscriptions are flat-rate; API pricing applies only if Claude Code runs against an API key. Subscribers meet pricing indirectly, as session and weekly usage limits rather than dollars.

Is Sonnet 5 more expensive than Sonnet 4.6 after September 1?

Per token, no: both list at $3/$15. Per unit of text, yes: Anthropic's docs state the newer tokenizer produces approximately 30 percent more tokens for the same text, so an identical workload meters roughly 1.0 to 1.35 times more on Sonnet 5.

Sources for this post: token counts come from my own ~/.claude/projects/ transcripts, 136 session files covering 2026-07-21 to 2026-08-04, deduplicated by message.id in a single pass on the morning of 2026-08-04; the session doing the measuring appends records as it runs, so exact counts drift slightly between passes. All rates, the August 31 cutoff, and the tokenizer note are quoted from Anthropic's pricing documentation and Sonnet 5 announcement, both read on 2026-08-04. Community quotes are from the linked Hacker News thread, read the same day. Dollar figures are notional API prices applied to a subscription workload that was never billed them.