← All articles

July 13, 2026

Claude Sonnet 5's Price Cut Didn't Actually Make Your AI Agents Cheaper

Claude Sonnet 5's per-token price didn't drop. A new tokenizer and default-on reasoning mean real AI agent costs can go up, not down.

ST

Stuart

Founder, Hotkey

Anthropic dropped Claude Sonnet 5 on June 30 and called it a cheaper way to run agents. Cheaper is the headline every outlet ran with. It’s also only half true.

If you’re running always-on AI agents for a sales team, or pricing a support retainer around one, the sticker price is the least useful number in this release. Two changes buried in the same announcement can push your actual bill up, not down, depending on what your agent is doing.

What Anthropic actually shipped on June 30

Claude Sonnet 5 launched at $2 per million input tokens and $10 per million output tokens. That rate holds through August 31, 2026. After that, pricing reverts to $3 and $15 per million tokens, the same per-token price as the previous model, Claude Sonnet 4.6. The real discount is temporary and only lasts nine weeks.

Anthropic says this plainly in its own documentation: per-token pricing on Sonnet 5 is unchanged from Sonnet 4.6. The “price cut” most coverage described is an introductory window, not a permanent repricing. Come September 1, you’re paying the same rate you were paying in June, for a different model with different behavior. It’s a different mechanism than the one we wrote about when HubSpot moved to pay-per-outcome pricing for its own AI agents, but the lesson is the same: read what’s actually changing before you read the headline number.

A hungrier tokenizer and thinking you didn’t ask for

Two changes shipped alongside Sonnet 5 that don’t show up in a pricing table but show up on your invoice.

First, Sonnet 5 uses a new tokenizer. The same input text now produces roughly 30% more tokens than it did on Sonnet 4.6, and the exact multiplier depends on content. Anthropic’s own numbers put it between 1.27x for code and 1.42x for English prose. You’re not sending more text. You’re just being charged for more of it.

Second, adaptive thinking is on by default. Sonnet 4.6 ran without extended reasoning unless you asked for it. Sonnet 5 decides for itself when a task needs deeper reasoning, and that reasoning burns output tokens even when you didn’t request it. You can turn it off by explicitly disabling thinking in the request, but out of the box, every call now includes reasoning tokens that weren’t there before.

Stack those two changes and a workload that used to cost a certain amount can end up costing meaningfully more, even though the per-token price on your invoice looks the same or lower.

What that actually does to the bill

Artificial Analysis ran Sonnet 5 through its Intelligence Index benchmark at standard pricing and found the average task cost $2.29. The same benchmark on Sonnet 4.6 cost about $1.20. On Opus 4.8, the flagship model Sonnet 5 is supposed to be the budget alternative to, it cost about $1.97. Sonnet 5 came in more expensive than the model it’s meant to undercut.

Model	Input $/M tokens	Output $/M tokens	Tokens for same text vs. Sonnet 4.6	Avg. task cost
Sonnet 4.6	$3	$15	1x (baseline)	about $1.20
Sonnet 5, introductory (through Aug 31, 2026)	$2	$10	about 1.3x	about $1.50 (estimated)
Sonnet 5, standard (from Sep 1, 2026)	$3	$15	about 1.3x	$2.29
Opus 4.8	$5	$25	n/a (different model)	about $1.97

During the introductory window, that same task would land somewhere around $1.50 (two-thirds of the standard-rate figure, since input and output pricing are both discounted by the same ratio). That’s cheaper than Sonnet 4.6, but by about 25%, not the 60% difference the per-token numbers suggest on their own. Wait until September, and the comparison flips: Sonnet 5 at standard pricing costs more per task than the model it replaced.

Where the discount is real

None of this makes Sonnet 5 a bad deal. It makes it a deal that depends entirely on what you’re asking it to do.

Short, structured tasks are where Sonnet 5 genuinely gets cheaper. Summarizing a call transcript into fixed CRM fields, drafting a templated follow-up email, classifying a lead against a fixed rubric. These are jobs where the input is short, the output is short, and there’s no real ambiguity that needs extra reasoning. Turn adaptive thinking off for tasks like this and the cost tracks close to the discounted per-token rate.

Prompt caching helps here too, and it has nothing to do with which model you pick. A cache hit costs 10% of the standard input price. An agent that re-reads the same CRM record structure or the same system prompt on every run should be caching that context regardless of model. That’s the biggest lever most teams aren’t pulling, and it matters more than the Sonnet 5 pricing story does.

Where it isn’t, and what changes on September 1

Open-ended agentic work is where the discount disappears. Multi-step account research, triaging an ambiguous deal, anything where the model has to reason through several possibilities before landing on an answer. That’s exactly the kind of task adaptive thinking was built for, and exactly where it adds the most tokens. On agent-heavy benchmarks, Sonnet 5 can run close to three times as many reasoning loops as Sonnet 4.6 to reach the same answer.

None of these numbers come from Hotkey’s own agents. I didn’t run our own Post-Call or Follow-Up systems through this comparison for this post, so treat the figures here as public and directional rather than a measurement of your specific workflow. The tokenizer multiplier alone ranges from 1.27x to 1.42x depending on the kind of text you’re feeding it, and your real cost will vary with your prompts, your context size, and how much of it is cacheable.

Don’t take “60% cheaper” as an input to your pricing or your architecture without checking your own workload first. If your agents run short, structured, low-reasoning tasks with cached context, Sonnet 5 during the introductory window is a real, if modest, saving. If they run open-ended reasoning work, you may already be paying more than you were in June, and September 1 makes it worse. Anyone quoting or running always-on agents for clients should recount their actual token usage against Sonnet 5’s tokenizer before repricing anything.

Frequently asked questions

Did Claude Sonnet 5 actually get cheaper than Claude Sonnet 4.6?

Only during the introductory window that runs through August 31, 2026, when it’s priced at $2 per million input tokens and $10 per million output tokens. Standard pricing from September 1 onward is $3 and $15 per million tokens, the same per-token rate as Sonnet 4.6. Because Sonnet 5’s tokenizer also produces about 30% more tokens for the same text, even the introductory window is a smaller real discount than the headline price suggests.

What is Claude Sonnet 5’s new tokenizer and why does it affect cost?

Sonnet 5 tokenizes text differently than Sonnet 4.6, and the same input now produces roughly 30% more tokens on average, ranging from about 1.27x for code to 1.42x for English prose. Since you’re billed per token, the same request can cost more even though the per-token price hasn’t changed.

What is adaptive thinking, and can I turn it off?

Adaptive thinking is Sonnet 5’s built-in reasoning mode, and it runs by default on every request, unlike Sonnet 4.6 where extended reasoning only ran if you asked for it. You can disable it explicitly in the API request, and doing so is worth testing on any short, structured task where the model doesn’t need extra reasoning to get the answer right.

Will always-on AI sales agents cost less to run with Sonnet 5?

It depends on the task. Short, structured jobs like call summaries or templated follow-ups, paired with prompt caching and thinking turned off, do get cheaper. Open-ended work like account research or deal triage can end up costing more than it did on Sonnet 4.6, because the tokenizer and default reasoning both add tokens exactly where that kind of task needs them.

What happens to Sonnet 5 pricing after August 31, 2026?

Standard pricing takes effect on September 1, 2026: $3 per million input tokens and $15 per million output tokens, the same rate as Sonnet 4.6. Combined with the new tokenizer’s higher token counts, a workload that looked cheaper in July can cost more than it did in June once the introductory rate expires.

— Stuart, Hotkey