What exactly changes on 1 September 2026?
Claude Sonnet 5's introductory API pricing of US$2 per million input tokens and US$10 per million output tokens runs through 31 August 2026. From 1 September it moves to standard pricing of US$3 per million input and US$15 per million output. That is a 50% increase on the headline rate, confirmed in Anthropic's own launch post.
Everyone is reading this as a 50% rise. For most practitioners running real workloads, the effective increase is larger than that, and the reason is buried in a footnote.
The extractable facts
--- Model: Claude Sonnet 5, released 30 June 2026, API name claude-sonnet-5.
--- Introductory price, through 31 August 2026: US$2 per million input tokens, US$10 per million output tokens.
--- Standard price, from 1 September 2026: US$3 per million input tokens, US$15 per million output tokens.
--- Headline increase: 50% on both input and output.
--- Opus 4.8, for reference: US$5 per million input tokens, US$25 per million output tokens.
--- Tokenizer note from Anthropic: the same input can map to roughly 1.0 to 1.35 times more tokens than on Sonnet 4.6.
--- Chat plan users are not billed per token and see no direct price change from this.
All of the above is stated on Anthropic's Claude Sonnet 5 announcement, including the availability and pricing section and its footnotes.
Why does the increase land harder than the headline 50%?
Sonnet 5 ships with an updated tokenizer. Anthropic states directly that the same input can map to roughly 1.0 to 1.35 times more tokens than before, depending on content type. Token count is the billing unit, so a 50% rate rise stacked on top of a possible 35% token inflation compounds rather than adds.
Anthropic was transparent about this. The introductory pricing was explicitly set so that moving from Sonnet 4.6 to Sonnet 5 would be roughly cost-neutral. The introductory period was, in effect, the cushion for the tokenizer change.
When the cushion is removed on 1 September, both changes land on your invoice at once.
Whether you feel the full 1.35 times or something closer to 1.0 depends on your content. Dense technical text, code, tables and Chinese-language input behave differently from plain English prose. There is no shortcut here. You have to measure your own workload.
That measurement is the single most valuable thing you can do in the next three weeks, and it takes about ten minutes.
How much will this actually cost you?
A workload consuming 20 million input tokens and 4 million output tokens per month costs US$80 at introductory pricing. From 1 September the same token volume costs US$120, a rise of US$40. If your tokenizer inflation runs at 1.25 times, the real figure is closer to US$150, roughly HK$1,170 per month.
These numbers assume no caching and no batching, which is exactly how most no-code automations are configured by default.
Worked examples at standard pricing
--- Light use, 5M input and 1M output per month: US$15 introductory, US$30 standard.
--- Moderate use, 20M input and 4M output per month: US$80 introductory, US$120 standard.
--- Heavy use, 100M input and 20M output per month: US$400 introductory, US$600 standard.
--- Add tokenizer inflation of 1.25 times on the input side and each standard figure rises by roughly a further 12 to 20%.
If you run Claude through a no-code platform such as Zapier, Make or n8n, or through a wrapper tool that resells API access, check whether your provider passes the change through on 1 September or absorbs it temporarily. Not all of them will tell you.
How do you cut the bill before 1 September?
Three levers do most of the work: prompt caching for repeated context, the Batch API for anything that tolerates delay, and effort level tuning so you stop paying premium reasoning rates for simple tasks. Anthropic documents prompt caching as cutting repeated input cost by up to 90%, and batch processing at 50% off.
The order matters. Measure first, then cache, then batch, then consider switching models. Switching first is how people end up with a cheaper bill and worse output.
Try this prompt
Paste this into Claude or ChatGPT along with a copy of one of your production prompts.
You are an LLM cost auditor. I am pasting a production prompt I run repeatedly. Analyse it and return the following, with no preamble.
1. Estimated input token count, and which parts of the prompt are identical on every single run.
2. A rewritten version that moves every unchanging part into a single stable block at the top, so it can be cached, and leaves only the variable parts at the bottom.
3. The three longest sections that could be shortened without changing the output quality, with a suggested replacement for each.
4. A yes or no answer to whether this task tolerates a delay of several hours, and therefore whether it is a batch candidate.
5. My estimated monthly cost at US$3 per million input tokens and US$15 per million output tokens, assuming [X] runs per month.
MY PROMPT: [paste it here]
The output of step 2 is usually where the money is. Most production prompts carry a large unchanging instruction block that is re-sent, and re-billed, on every call.
Should you switch models, and to what?
Only if you have measured first. At US$3 and US$15 Sonnet 5 still sits well below Opus 4.8 at US$5 and US$25, and industry comparison tables published in 2026 place Gemini 3.1 Pro at roughly US$2 and US$12 and GPT-5.6 Sol at roughly US$5 and US$30. Cheaper models exist, and they cost more when a task has to be retried three times.
Be honest about where the competition wins. Gemini 3.1 Pro is the cheaper option on the input side for large-context work. Budget models such as the DeepSeek and Amazon Nova families are dramatically cheaper per token and are a genuine fit for high-volume, low-judgement tasks like classification and tagging.
Where Sonnet 5 holds up is sustained agentic work: multi-step tasks with tool use where a model that quits halfway costs you more than its token price ever saved.
A decision rule that survives contact with reality
--- Classification, tagging, extraction at volume: use a budget model. The task has a checkable right answer.
--- Drafting and editing where a human reviews the output anyway: mid-tier is fine. Test two models on ten real jobs.
--- Multi-step automation with tool calls and no human in the loop: pay for the stronger model. Failure is more expensive than tokens.
--- Anything you cannot evaluate: do not switch. You will not notice the quality drop until a client does.
Prices quoted for third-party models come from published 2026 comparison tables rather than each vendor's own page, so verify against the vendor before you commit.
What does none of this solve?
Cost optimisation does not fix a workflow that was never worth running. If an automation produces output nobody uses, caching it makes it 90% cheaper to waste. The uncomfortable question before any pricing exercise is which of your AI workflows actually changed a decision or saved a person time last month.
There are also honest limits to the alternatives, including ours.
Where the alternatives fall short
--- Prompt caching only pays off when the cached block is genuinely repeated and stable. Change it often and you pay a write premium for nothing.
--- The Batch API is unusable for anything a person is waiting on. Customer replies and live research do not qualify.
--- Budget models degrade unevenly. They rarely fail loudly. They fail in the 5% of cases that matter most.
--- UD's AI Employee Hub is free at HK$0 per month and covers 8 prebuilt roles with 24 skills, but it is prebuilt. If your workflow needs bespoke logic against your own data, it is a starting point, not a replacement for building something.
--- Nothing here removes the need to evaluate output quality yourself. No pricing page can do that for you.
Our earlier piece comparing what US$20 a month actually buys across ChatGPT, Claude and Gemini covers the subscription side of this question, which is a different decision from per-token API billing.
What is the correct next step before 31 August?
Do three things this week. Pull your actual token usage for July from your provider console. Run the audit prompt above on your two highest-volume prompts. Then model September at US$3 and US$15 and decide, with a number in front of you, whether to optimise, switch, or accept the increase.
You have until 31 August. That is roughly three weeks from now, and the work is a single afternoon.
The teams that will be surprised in September are the ones treating this as a headline rather than a line item. The ones who measure will find the increase was never the real problem; the unexamined workflow was.
Cost pressure has a way of making technology feel cold. It does not have to. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Want a Second Opinion on Your AI Bill Before 1 September?
If your token bill is about to rise and you are not sure which workflows are worth keeping, start with the free option. UD's AI Employee Hub gives Hong Kong businesses 8 prebuilt AI employees with 24 skills at HK$0 per month, no salary and no onboarding. And we'll walk you through every step, from measuring your current usage to deciding what belongs on an API and what does not.
