What does one AI task actually cost in 2026?
Summarising a 40-page report costs between US$0.02 and US$0.13 depending on which model you send it to, which is roughly HK$0.15 to HK$1.00. The spread between the cheapest and the most expensive option for the same job is about seven times, and for most routine work the expensive option is not seven times better.
That number surprises people who only know AI through a monthly subscription, because a subscription hides per-task cost completely. The moment you run AI inside an automation, a spreadsheet script, or a tool like Make or n8n, you start paying per token and the arithmetic becomes visible.
This page is the arithmetic. Real published prices as of 19 August 2026, one worked example, and an honest account of when these numbers stop being true.
What do GPT-5.6, Gemini 3.7 Flash and Claude Sonnet 5 cost per million tokens?
Published API prices as of 19 August 2026, input and output per one million tokens. A million tokens is roughly 750,000 English words. Prices below come from each vendor's own announcement, not from a third-party tracker.
Current published prices
--- Gemini 3.7 Flash: US$0.75 input / US$3.75 output. Introductory pricing announced 13 August 2026, valid to 31 December 2026. From 1 January 2027 it becomes US$1.50 / US$7.50, per Google's own footnote.
--- GPT-5.6 Luna: US$1 input / US$6 output. OpenAI's fast, lowest-cost tier.
--- Claude Sonnet 5: US$2 input / US$10 output. Anthropic announced on 10 August 2026 that this introductory price is now permanent, cancelling the US$3 / US$15 rate that was due to start on 1 September.
--- GPT-5.6 Terra: US$2.50 input / US$15 output. OpenAI's balanced everyday tier, described as competitive with GPT-5.5 at half the price.
--- GPT-5.6 Sol: US$5 input / US$30 output. The flagship tier, built for long-horizon agentic work.
Two of these five prices carry an expiry date. That is the single most important fact on this page for anyone building a workflow they intend to still be running in 2027.
How much does one real task cost: summarising a 40-page report?
A 40-page report is roughly 15,000 words, which is about 20,000 input tokens. An 800-word summary is about 1,100 output tokens. Multiply those through each price and you get the true cost of one run.
Cost of one 40-page summary
--- Gemini 3.7 Flash: US$0.019, about HK$0.15
--- GPT-5.6 Luna: US$0.027, about HK$0.21
--- Claude Sonnet 5: US$0.051, about HK$0.40
--- GPT-5.6 Terra: US$0.067, about HK$0.52
--- GPT-5.6 Sol: US$0.133, about HK$1.04
Hong Kong dollar figures use an approximate 7.8 conversion and exclude payment processing and any platform fee your automation tool adds on top.
Note what dominates. Input is 20,000 tokens and output is 1,100, so on a summarisation job the input price drives the bill. On a job that generates long output, a marketing brief expanded into ten pages, the output price drives it instead, and the ratios change completely.
At what volume does an API bill beat a monthly subscription?
At 100 report summaries a month, the API cost ranges from about US$1.91 on Gemini 3.7 Flash to US$13.30 on GPT-5.6 Sol, which is roughly HK$15 to HK$104. A single US$20 consumer subscription seat costs about HK$156 a month regardless of whether you run one task or 400.
100 summaries a month, by model
--- Gemini 3.7 Flash: about US$1.91, HK$15
--- GPT-5.6 Luna: about US$2.66, HK$21
--- Claude Sonnet 5: about US$5.10, HK$40
--- GPT-5.6 Terra: about US$6.65, HK$52
--- GPT-5.6 Sol: about US$13.30, HK$104
The honest reading is not "cancel your subscription". A subscription buys the app, file handling, image generation, voice, and a fixed bill you cannot accidentally blow up at 2am. The API buys per-task economics and the ability to put the model inside a workflow that runs without you.
The decision point is whether the work is repetitive and scheduled. One-off thinking belongs in the chat app you already pay for. The same task 100 times a month belongs in an automation billed per token.
How does prompt caching change the bill?
Prompt caching bills repeated context at a heavy discount, and for workflows that resend the same brand guide, price list or system prompt on every run, it is usually the largest single saving available. OpenAI states that for GPT-5.6 and later, cache reads receive a 90% discount on the input rate while cache writes are billed at 1.25 times the uncached input rate.
GPT-5.6 also introduced explicit cache breakpoints and a 30-minute minimum cache life. The practical consequence is that a workflow firing every 20 minutes keeps its cache warm, and one firing every hour pays the write premium again on each run.
The design rule that follows is simple. Put the stable material first in the prompt, the brand guide, the examples, the rules, and the variable material last. Caching works on the unchanged prefix, so reordering a prompt can cut the input bill without changing a word of the content.
A worked case: a 4,000-token style guide resent 500 times a month on GPT-5.6 Luna costs US$2 uncached and about US$0.20 on cache reads. The saving is small in absolute terms and large as a proportion, which is exactly how these bills grow unnoticed.
Which model should you route each job to?
Route by consequence of failure, not by prestige. Classification, tagging, routine summaries and first-pass drafts go to the cheapest capable tier. Anything a client reads unedited, or anything that runs unattended across many steps, justifies the flagship price.
Verdict by user type
--- Solo marketer or freelancer running scheduled content jobs: Gemini 3.7 Flash or GPT-5.6 Luna. At these volumes the bill stays under HK$50 a month and the quality gap on routine text is small.
--- Anyone whose output is long-form writing read by clients: Claude Sonnet 5. Its price sits mid-table and it is the model practitioners most consistently name for prose that needs less cleanup.
--- Operations lead automating multi-step business workflows: GPT-5.6 Terra as the default, escalating to Sol only for the steps where a wrong answer is expensive.
--- Anyone processing complex PDFs and structured documents: Gemini 3.7 Flash. Google reports it scoring 34.0% versus 3.6 Flash's 22.0% on the GDP.pdf document-comprehension benchmark.
--- Long-horizon agent work across many tool calls: GPT-5.6 Sol. It is the tier OpenAI built for persistence, and the only one where the premium is defensible on a routine job.
Where do these numbers stop being true?
Every figure on this page is a list price for text tokens on a standard tier, and several real costs sit outside that boundary. Treat the table as a floor, not a forecast.
--- Two of the five prices expire. Gemini 3.7 Flash doubles on 1 January 2027 by Google's own published schedule. Claude Sonnet 5's rate is now permanent, but "permanent" is a company statement, not a contract.
--- Reasoning tokens are billed as output. A model set to high or maximum reasoning effort can generate several times more billed tokens than the visible answer suggests. GPT-5.6 introduced a new maximum reasoning setting, and it is not free.
--- Images, audio and video bill differently. Nothing here applies to image generation or multimodal input, which use separate rates and often per-item pricing.
--- Your automation platform charges on top. Make, Zapier and n8n bill per operation or per execution regardless of the token cost underneath. On small tasks that platform fee frequently exceeds the model fee.
--- Consumer subscription tiers are a different product. Rate limits, model access and features on a chat plan are not the same as API access, and you cannot read one price off the other.
What if you do not want to manage models at all?
Then the per-token table is the wrong tool for you, and the honest alternative is a packaged setup where someone else has already chosen the model. UD's AI Employee Hub is one such option: 8 pre-built AI employees covering 24 skills, at HK$0 a month, provided free to Hong Kong businesses.
The roles are specific rather than generic. Amy handles social media, David covers SEO and content, Eva does scheduling, documents and meeting notes, Frank builds data dashboards, and Henry runs security checks including PDPO compliance and vulnerability scanning.
Be clear about the trade-off. The Hub does not expose model selection, does not give you a per-token bill to optimise, and will not help you decide between Luna and Terra. If routing by cost is your actual problem, the Hub does not solve it and this page's table does.
It solves the opposite problem: you want the output, not the infrastructure decision. If you have spent two evenings comparing token prices for work that runs twice a week, that is the signal you are optimising the wrong layer.
The short version
One AI task in 2026 costs cents, not dollars, and the difference between models on routine work is measured in tens of Hong Kong dollars a month rather than hundreds. The expensive mistakes are structural: an uncached system prompt resent thousands of times, a reasoning setting left on maximum, or a flagship model doing classification work.
Check the two prices with expiry dates before you build anything you expect to run next year. Then route by consequence, cache the stable material, and stop paying flagship rates for tasks that never needed them.
If you want to go one level deeper on output quality rather than cost, our guide to chain-of-verification prompting covers the four-minute loop that catches the facts a cheaper model gets wrong. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Skip the Model Comparison Entirely
If you would rather have working AI staff than a spreadsheet of token prices, start with 8 ready-made AI employees at HK$0 a month. We'll walk you through every step, from choosing the right role to installing it in your existing tools.
