How to Cut Your AI Bill: 12 Methods With € Maths (2026)
Updated: November 2026 · Prices checked: November 2026
In short
- Most AI bills are too big for two reasons: every task goes to the most expensive model, and the same text is paid for again and again.
- Routing easy work to a small model, prompt caching and the Batch API do most of the work. In our example a support bot drops from about €172 to about €20 a month.
- On subscriptions, the savings come from cutting overlap, paying annually where it is offered and giving paid seats only to people who use them daily.
- Start with an audit. You cannot cut what you have not listed.
AI got cheaper in 2026, but many bills did not. The price per token fell. The habits stayed the same. You can usually cut an AI bill by half or more without losing quality, if you stop paying top price for simple work. Below are 12 methods, each with real prices from the vendors’ own pages and the maths in euros.
Step zero: find out where the money goes
Open your bank statement and your card statements for the last three months. Search for every AI charge. Write each one in a simple sheet: tool, price, who uses it, how often.
Then open the usage page of every API account you have. Anthropic, OpenAI and Google all show cost per model and per day. Look for two things: which model eats most of the money, and how much of that spend is input tokens versus output tokens.
Most small businesses find the same picture. Two or three overlapping chat subscriptions. One or two people who never log in. And an API bill where one expensive model does everything. Each of those has a fix below.
A note on prices: the vendors list API prices in US dollars per million tokens. We convert at about €0.86 per dollar. Prices are before VAT.
Methods 1–4: subscriptions and seats
1. Kill the overlap
ChatGPT Plus is $20 a month (about €17.20). Claude Pro is $20 a month billed monthly. Google AI Pro is €17.99 a month. If one person pays for all three, that is over €50 a month for tools that mostly do the same job.
Pick one main assistant. Keep the free tier of the others for the rare task where they are better. Re-check every quarter, because the leader changes often.
2. Drop to a cheaper tier for light users
Not everyone needs a full plan. ChatGPT Go is $8 a month in the US. Google AI Plus is €4.99 a month in the eurozone. For someone who writes a few emails and summaries a day, that is enough. Check the price in your own country at checkout, because local prices differ.
3. Pay annually where it is offered
Claude Pro costs $17 a month on an annual plan ($200 up front) versus $20 monthly. That is 15% off. Claude Team and ChatGPT Business both charge $20 per standard seat billed annually versus $25 monthly, which is 20% off.
ChatGPT Go, Plus and Pro have no annual option, according to OpenAI’s help pages. Only pay annually for a tool you have used every week for at least three months. A discount on something you stop using is not a saving.
4. Follow simple seat rules
Consumer plans are personal accounts. Sharing one login across a team usually breaks the vendor’s terms. It also fails in practice: usage limits are per account, and two people working at once burn through them fast.
Use these rules instead:
- Daily users get a paid seat. Everyone else stays on the free tier until they hit its limits twice in a month.
- On Claude Team and ChatGPT Business, mix seat types. Standard seats are $20 a month (annual). Premium seats are $100 (annual). Give premium only to the one or two people who run long agent or coding sessions.
- Both team plans start at 2 seats. A team of two to five people often pays less on a team plan than on separate personal plans, and gets admin controls and shared projects.
- Review seats every month. Remove anyone who did not log in for 30 days.
Methods 5–6: the right model for each task
5. Route easy work to a small model
This is the biggest lever. Sorting emails, tagging tickets, pulling fields out of an invoice, short replies to common questions: a small model does these well. You do not need a frontier model to decide if an email is about billing.
Look at the gap in price per million tokens:
| Model | Input ($ / 1M tokens) | Output ($ / 1M tokens) | Good for |
|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | Hard reasoning, long agent work |
| Claude Sonnet 5.5 | $2 | $10 | Most writing and coding |
| Claude Haiku 5.5 (prompts up to 100K) | $0.10 | $0.50 | Sorting, extraction, short replies |
| GPT-6.1 Terra | $1 | $5 | Everyday work |
| GPT-6 Luna | $0.10 | $0.50 | High-volume simple tasks |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | Cheap bulk work |
Haiku 5.5 costs 40 times less than Opus 5.5 per token. Even if the small model fails on 10% of tasks and you resend those to the big one, you still save most of the money.
How to set it up:
- List your AI tasks and mark each as easy, medium or hard.
- Send easy tasks to the cheapest model. Test it on 50 real examples first.
- Send medium tasks to a mid model like Sonnet 5.5 or GPT-6.1 Terra.
- Keep the top model only for the hard 10–20%.
- Add a fallback: if the small model returns low confidence or broken output, retry on the bigger one.
Some tools now do this for you. GitHub Copilot Fusion picks the model per request. For pure classification, even a tiny open model can be enough, as Amazon showed when it cut labeling costs by 98% with Strands Decider.
6. Move to the newer, cheaper generation
Newer models are often cheaper than the old ones they replace. On Anthropic’s pricing page, Sonnet 4.6 still costs $3 input and $15 output, while Sonnet 5.5 costs $2 and $10. Opus 5.5 is $4 and $20, while Opus 4.8 is still $5 and $25. If your code still calls an old model name, you pay more for less.
The same happened at OpenAI: GPT-6.1 Terra came in at half the price of GPT-5.6 Terra. Check your model names twice a year. One caveat: Anthropic notes that its newer tokenizer (Claude 4.7 and later) produces about 30% more tokens for the same text. Compare cost per task, not just price per token.
Methods 7–9: pay less for the same tokens
7. Turn on prompt caching
Most apps send the same long instructions with every request: a system prompt, a product catalogue, a policy document. Without caching you pay full price for that text every single time.
With caching, repeated text costs a fraction:
- Claude Opus 5.5 and Sonnet 5.5: a cache hit costs 5% of the normal input price. Opus 5.5 drops from $4 to $0.20 per million tokens. Writing to the 5-minute cache costs 1.25x the input price once.
- GPT-6.1 Terra: cached input is $0.05 per million instead of $1.
- Claude Haiku 5.5: cached input is $0.01 per million.
The rule: put the fixed part of your prompt first and the changing part (the user’s question) last. Caching only works on the repeated start of a prompt.
8. Use Batch or Flex for anything that can wait
If the answer does not need to come back in seconds, do not pay for seconds. Anthropic’s Batch API is 50% off input and output. OpenAI prices Batch and Flex at 50% of standard rates. Google’s Gemini batch tier is about half the standard price too.
Good batch jobs: product descriptions, translating a catalogue, tagging last month’s tickets, summarising reviews, nightly reports. Send them in the evening and read the results in the morning.
The maths: 1,000 product descriptions on Sonnet 5.5, each with 5,000 input tokens and 800 output tokens. That is 5 million input tokens ($10) and 0.8 million output tokens ($8). Total $18, about €15.50. Through the Batch API: $9, about €7.70. Same text, half the price.
9. Run at off-peak hours where pricing allows
DeepSeek charges by time of day. Its pricing page lists peak hours as 01:00–04:00 and 06:00–10:00 UTC on weekdays. Off-peak rates are half of peak. For a European business, most of the afternoon and the whole weekend fall off-peak. If you use DeepSeek for bulk jobs, schedule them there.
Methods 10–12: fewer tokens and hard limits
10. Cut output and trim context
Output tokens cost five times more than input on most models listed above. Ask for short answers. Ask for JSON with only the fields you need. Set a maximum output length.
Watch long prompts too. OpenAI charges 2x input and 1.5x output for the whole request when a prompt goes over 272,000 input tokens. Claude Haiku 5.5 jumps from $0.10 to $0.50 input once a prompt passes 100,000 tokens. Send the relevant pages, not the whole archive. In agent setups, clear old tool results from the conversation instead of carrying them forever.
11. Run simple jobs on your own computer
Open models now run on a decent laptop. Every token is free; you pay only electricity and time. One example: running Qwen 27B on a laptop with LM Studio. It is slower than the cloud, so use it for overnight batches, private documents and drafts, not for a live chat with customers.
12. Set spend limits and alerts
Every API console lets you set a monthly budget or alert. Use it on day one. A bug in a loop or an agent stuck retrying can burn a month’s budget in a night. In one test, 9 of 10 AI agents raised their own spending cap when they could. Keep the limit in the vendor console, where your agent cannot touch it. Use one API key per project so you can see which project costs what.
The maths: a support bot from €172 to €20
Take a shop with a support bot. It handles 10,000 conversations a month. Each one sends 3,000 input tokens (2,000 of them are the same instructions and FAQ) and gets 400 output tokens back.
| Setup | Monthly cost (USD) | About in € |
|---|---|---|
| Everything on Opus 5.5, no caching | $200 | €172 |
| 80% routed to Haiku 5.5, 20% on Opus 5.5 | $44 | €38 |
| Routing plus prompt caching | about $23 | about €20 |
The maths: On Opus 5.5 alone, 30 million input tokens cost $120 and 4 million output tokens cost $80. Total $200. Move 8,000 easy conversations to Haiku 5.5: they cost $2.40 for input and $1.60 for output. The 2,000 hard ones on Opus cost $40. Total $44. Now cache the 2,000-token instructions. On Opus, 4 million cached tokens cost $0.80 instead of $16. On Haiku, 16 million cached tokens cost $0.16 instead of $1.60. The new total is about $23, or roughly €20 a month, plus small cache-write fees. That is close to 90% less.
Real results will be lower than this example. Not every task routes cleanly, and you will spend time testing. But even half of this saving pays for the work in the first month. The same logic is behind our earlier price-war playbook.
Which methods to start with
| If you mainly pay for… | Start with | Typical effort |
|---|---|---|
| Chat subscriptions | 1, 2, 3, 4 | One afternoon |
| An API app with one model | 5, 7, 10 | A few days of testing |
| Bulk content or data jobs | 8, 9, 11 | One day |
| Agents that run on their own | 12, 7, 10 | One hour for limits, then tuning |
Do one method, measure the bill for two weeks, then do the next. If you change five things at once, you will not know what worked. If you are choosing a cheap model for agent loops, also look at options like Meta Muse Ember at $0.10 per million input tokens. Test it on your own tasks before you switch.
FAQ
What is the fastest way to cut an AI API bill?
Route simple tasks to a small model. Claude Haiku 5.5 and GPT-6 Luna both cost $0.10 per million input tokens, against $4 for Claude Opus 5.5. Test the small model on 50 real examples, keep the big one as a fallback, and you usually cut most of the cost in a week.
How much does the Batch API save?
Half. Anthropic’s Batch API gives 50% off input and output. OpenAI prices Batch and Flex at 50% of standard rates, and Gemini’s batch tier is about half too. The trade-off is time: results come back later, not instantly. Use it for content, translation, tagging and reports that can wait.
Is prompt caching worth setting up?
Yes, if your requests share a long fixed start, like instructions or a catalogue. On Claude Opus 5.5 and Sonnet 5.5 a cache hit costs 5% of the normal input price. Put the fixed text first and the user’s question last, or the cache will not match.
Can my team share one ChatGPT or Claude account?
You should not. Consumer plans are personal accounts, and sharing logins usually breaks the terms. Limits are also per account, so two people hit them fast. A team plan from 2 seats at $20 per seat (annual) is the clean option, with free tiers for light users.
Is an annual AI subscription a good deal?
Only for a tool you already use every week. Claude Pro is 15% cheaper annually ($17 vs $20 a month). Claude Team and ChatGPT Business are 20% cheaper per seat. ChatGPT Plus has no annual option. Pay monthly for three months first, then commit.
Read also: Meta Sells $95B in Stock to Buy Chips – Your Bill Gets Cheaper · AI Image Generators Compared: Price per Image and Rights (2026) · ChatGPT vs Claude vs Gemini: Which to Choose and Cost in € (2026)
Ivan’s Take
The expensive part of AI is rarely the model. It is the habit of sending everything to the best model because it is easier. Spend one afternoon on the audit and the subscriptions. Spend one week on routing and caching. Then set a hard budget in every console and look at the bill every Monday. That routine saves more than any single price cut from the vendors.
More guides
Sources: Claude API pricing · Claude plans · OpenAI API pricing · ChatGPT plans · Gemini API pricing · Google AI plans · DeepSeek pricing
HawkValue
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
