In this guide · 8 sections
- 01Two discounts most AI projects leave unused
- 02Published caching and batch discounts, October 2026
- 03How prompt caching works, in plain terms
- 04Batch APIs: half price for work that can wait
- 05Worked example: a website chatbot and an overnight job
- 06Other ways to cut the bill
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Prompt caching makes repeated prompt text (instructions, FAQ) cost 5% to 10% of the normal input price after the first write. Batch APIs cost 50% less at OpenAI, Anthropic and Google for jobs that can wait up to 24 hours. Use caching for chatbots and batch for overnight jobs.
Two discounts most AI projects leave unused
AI APIs charge per token, for what you send in and what comes back. Our guide to LLM API pricing per token explains the basics. Two vendor features change the maths:
Prompt caching
For live chat- The repeated start of your prompt is stored briefly
- Reusing it costs a fraction of normal input
- Answers are as fast or faster
- Works best when the same instructions are sent many times an hour
Best for: Chatbots, assistants and agents with long fixed instructions.
Batch API
For work that can wait- Send many requests in one file
- Results come back later, up to 24 hours at OpenAI
- 50% off input and output
- No good for anything a customer is waiting for
Best for: Overnight jobs: product descriptions, summaries, tagging.
Published caching and batch discounts, October 2026
From OpenAI's API pricing, Anthropic's pricing page and Google's Gemini API pricing, in US dollars per million tokens, checked on 22 October 2026. Approximate; vendors change prices often.
| Model | Normal input | Cache write | Cached input | Batch input / output |
|---|---|---|---|---|
| GPT-6.1 Sol (OpenAI) | $2.00 | $2.50 (30 min) | $0.10 | $1.00 / $5.00 |
| GPT-6 Luna (OpenAI) | $0.10 | $0.125 (30 min) | $0.01 | $0.05 / $0.25 |
| Claude Sonnet 5.5 (Anthropic) | $2.00 | $2.50 (5 min) | $0.10 | 50% off |
| Claude Haiku 5.5 (Anthropic) | $0.10 | $0.125 (5 min) | $0.01 | 50% off |
| Gemini 4 Argon (Google) | $2.00 to 30 Nov, then $4.00 | Implicit only | $0.10, then $0.20 | $1.00 / $5.00 to 30 Nov |
| Gemini 3.8 Flash (Google) | $0.75 to 31 Dec, then $1.50 | Plus storage per hour | $0.075, then $0.15 | $0.375 / $1.875 to 31 Dec |
OpenAI also offers a 12-hour cache lifetime at 2.5 times normal input, and Anthropic a 1-hour lifetime at US$4.00 per million for Sonnet 5.5. Google's Gemini 4 Argon supports only implicit caching; Gemini 3.8 Flash also lists a cache storage charge of US$0.50 per million tokens per hour.
How prompt caching works, in plain terms
Every chatbot request resends the same instructions, product list or FAQ before the customer's question. With caching, the vendor keeps that repeated beginning for a short time and charges a small fraction to reuse it.
- It caches the start of the prompt only. Put fixed content first (instructions, FAQ, tool definitions) and the changing part (the customer's question) last. Anthropic says caching covers tools, system and messages in that order, up to the marked point.
- There is a minimum size. OpenAI's prompt caching guide says 1,024 tokens for GPT-5.6 and later; Anthropic lists 512 tokens for its current models; Google lists 4,096 tokens for Gemini 4 Argon and 3.8 Flash.
- It expires. OpenAI's default minimum lifetime is 30 minutes. Anthropic's is 5 minutes, refreshed free each time it is used. Quiet periods mean paying the write price again.
- The first write costs extra. OpenAI and Anthropic charge 1.25 times normal input for short-lifetime writes, so caching only pays if the same prefix is reused.
Batch APIs: half price for work that can wait
OpenAI's Batch API guide describes a 50% discount, a separate pool of higher rate limits and completion within 24 hours. Anthropic's batch processing gives 50% off input and output, and Google lists its Batch API at a 50% cost reduction.
| Good fit for batch | Not a fit |
|---|---|
| Rewriting 2,000 product descriptions overnight | A customer chatting on your website |
| Summarising last month's support emails | Replying to a WhatsApp enquiry |
| Tagging or classifying a backlog of documents | Anything a staff member is waiting on |
| Translating a catalogue or help articles | Real-time fraud or order checks |
Batch and caching can be combined: OpenAI lists batch prices for cached input too, and Anthropic says caching multipliers stack with the batch discount.
Worked example: a website chatbot and an overnight job
Illustrative arithmetic using the October 2026 list prices for GPT-6.1 Sol and Claude Sonnet 5.5, which are the same for these figures (US$2 input, US$10 output, US$2.50 cache write, US$0.10 cached input per million tokens).
- Chatbot, no caching3,000 chats x (3,200 input + 300 output tokens)about US$28.20/mo
- Chatbot, with caching3,000-token fixed prompt cached; assumes 90% cache hits, 10% re-writesabout US$13.26/mo
- Product rewrite, standard API2,000 products x (800 input + 400 output tokens)about US$11.20
- Product rewrite, batch APISame job at 50% off, results within 24 hoursabout US$5.60
Saving on both jobsabout US$20.54
Our arithmetic from list prices, before tax. Real cache hit rates depend on traffic: few chats an hour means more cache re-writes.
At small-business volumes the dollar amounts are modest, but the percentage saving holds as volumes grow. Our guide to what each customer reply costs in tokens shows how to estimate your own token counts.
Other ways to cut the bill
Caching and batch come after two bigger levers, plus one safety net:
Use the smallest model that passes your test
Small models like GPT-6 Luna or Claude Haiku 5.5 cost about a twentieth of mid-range models per token. Test them on your real questions first.
Trim what you send
Send only the relevant FAQ sections rather than the whole document, and cap output length.
Turn on caching
Fixed content first, variable content last; check cached tokens in the response.
Move waiting work to batch
Anything scheduled or overnight goes through the Batch API.
Set spending limits and alerts
Budgets and per-key limits stop a bug or abuse from running up a bill. See our guide to AI API cost control.
Do it yourself or get help?
If you call the APIs directly, caching is mostly about ordering your prompt and checking the usage numbers, and batch is a separate endpoint with a file upload. Many no-code chatbot tools don't expose either, so check before assuming you get the discount.
Our AI Website Assistant (S$790) builds a chatbot trained on your website, FAQ and documents, with guard-rails, lead capture and hand-over to WhatsApp or email. For batch jobs such as catalogue rewrites or document summaries, our Custom Tool or Dashboard starts from S$1,990 with a free written scope; API fees are paid by you in your own account.
Frequently asked questions
A feature of AI APIs that stores the repeated beginning of your prompt, such as instructions and FAQ text, for a short time so later requests can reuse it at a much lower price. In October 2026, cached input on GPT-6.1 Sol and Claude Sonnet 5.5 costs US$0.10 per million tokens versus US$2.00 normally, after a slightly dearer first write.
OpenAI, Anthropic and Google all list a 50% discount for batch processing. You send many requests together and receive results later; OpenAI says each batch completes within 24 hours. It suits overnight or scheduled work such as product descriptions, summaries and tagging, not live conversations with customers.
Usually because the start of the prompt changes between requests, for example a timestamp, the customer's name or a reordered list placed before the fixed instructions. Also check the minimum length: 1,024 tokens for recent OpenAI models, 512 for current Claude models and 4,096 for current Gemini models. Look for cached token counts in the API response.
No. Caching reuses the processed prompt rather than changing it, so the model sees the same input and answers the same way. It can also reduce response time. The only behaviour to plan for is that caches expire after a short period, so occasional requests pay the higher cache-write price again.
Yes, with the major vendors. OpenAI lists separate batch prices for cached input, and Anthropic says its caching multipliers stack with the batch discount. In practice the combination matters most for large batch jobs that share a long fixed instruction block, such as rewriting thousands of product descriptions in the same style.