If you build with AI, the model you pick can mean a 10x difference in your invoice. Here's who's cheapest per million tokens — and the 7 tricks that cut most bills 60-90%.
Across the major LLM APIs in 2026, per-million-token prices range from roughly $0.15 to $75 for input, and output tokens typically cost 3-5x more than input. At the cheap end, DeepSeek is consistently 10-50x less expensive than top Western models. The biggest wins are rarely the headline rate — they're prompt caching, batch APIs, and controlling output length. Apply all seven tricks and most AI bills drop by 60-90%.
A token is the chunk of text a model processes. As a rough rule of thumb: 1 token ≈ 4 characters ≈ three-quarters of an English word. So 1 million tokens is roughly 750,000 English words — a whole stack of books. Providers bill per million tokens, and they bill input (your prompt) and output (the model's reply) separately.
Almost every provider charges output tokens 3-5x their input price. The reason is technical: a model reads your input in parallel (fast, cheap), but it writes output one token at a time (sequential decoding — slower, compute-heavy). This single fact drives the single most effective saving below: keep outputs short.
Exact figures change week to week, so we group vendors into price tiers (verified against vendor pricing pages and independent trackers as of August 2026). Always confirm the exact number on the vendor's own pricing page before committing — headline rates also hide differences like tokenizers and minimum charges.
| Price tier | Vendor / model (example) | Ballpark input / output (per 1M tokens) | Cheap vs. top? |
|---|---|---|---|
| CHEAPEST | DeepSeek V3/V4, Gemini Flash, Qwen (small), Llama (small) | ~$0.10 - $0.60 in / ~$0.30 - $2 out | 10-50x cheaper |
| BUDGET | GPT mini-type, Gemini Flash-pro, Mistral small | ~$0.15 - $1 in / ~$0.60 - $4 out | 5-25x cheaper |
| MID | GPT-4o-class, Gemini Pro, Claude Sonnet | ~$1 - $3 in / ~$5 - $15 out | baseline mainstream |
| PREMIUM | Claude Opus, GPT flagship, Gemini Ultra | ~$5 - $15 in / ~$20 - $75 out | most expensive |
To make this concrete, here are real-world monthly estimates from independent trackers — cost to process 10 million tokens/month (a moderately busy product):
The takeaway isn't a single "best" number — it's that the spread between cheapest and most expensive for the same task is routinely 10x. Choose the cheapest model that still performs well enough for each specific job.
These are the high-leverage savings that don't sacrifice quality. Most teams stack several of them.
Most providers (Anthropic, OpenAI, Gemini, DeepSeek) offer a cached-input price that's dramatically cheaper — often a 90% discount — when you send the same long prefix (system prompt, context, documents) repeatedly. Put your stable context first and don't change it.
OpenAI, Anthropic and Gemini all offer batch / async endpoints at roughly half price when you can wait minutes-to-hours for results. Perfect for overnight processing, backfills, and evaluation runs.
Don't use the flagship for everything. Use a cheap model (DeepSeek, Gemini Flash, GPT mini) for classification, summarization, extraction and drafts; reserve premium models for the hard 10% of tasks.
Set a max_tokens cap and ask for concise output ("reply in under 50 words", "return JSON only"). Fewer output tokens = direct savings on the most expensive part of every call.
Remove redundant context, compress documents before sending, and don't re-send the same instructions every turn. Every input token you don't send is money you don't spend.
Most families ship a small + large pair (GPT-4o + GPT-4o mini, Gemini + Flash, Claude + Haiku). The mini often handles 80% of tasks at a fraction of the cost — use it first.
Run Llama, Qwen, Mistral or DeepSeek via a low-cost inference host or your own hardware. You trade setup effort for dramatically lower per-token costs — worth it at scale.
DeepSeek is consistently the cheapest major LLM API, often 10-50x cheaper than top Western models. Google Gemini Flash and OpenAI's mini models are also low-cost options.
Output tokens are generated one at a time (sequential decoding), which is more compute-intensive than reading input tokens in parallel. Most providers charge output 3-5x their input price. This is why controlling output length saves so much.
Use prompt caching (cached input is often 90% cheaper), use batch APIs (up to 50% off), pick cheaper models for routine tasks, and keep prompts and outputs short.
No. A cheap model that produces wrong answers or needs many retries can cost more in total than a pricier model that gets it right the first time. Match the model to the task's difficulty.
Prompt caching lets you send a long, stable prefix (system prompt, context) once and be billed at a much lower rate for it on subsequent calls. It's the single biggest saving for apps that reuse context.
Combine multiple PDF files into one — free, no upload, files stay on your device.
📒 Bookkeeping ReviewsHonest accounting software reviews: Wave, QuickBooks, Xero & free options.
🎮 Gaming Gear ReviewsEsports hardware decoded: gaming mice, keyboards, headsets & budget picks.
📷 ID Photo StudioPassport & visa photos with AI background changer, compressor & converter.
🧮 HK Salary Tax CalculatorEstimate Hong Kong salaries tax for 2026/27 in seconds.
🎁 HK Benefits Quick CheckGovernment & benefits eligibility checker — OAA, CSSA, housing & more.
📅 HK Calendar 香港月曆Public holidays, lunar dates & solar terms for 2020-2036 — bilingual, free, no sign-up.