Every major model provider prices API access the same fundamental way: by the token. But most people building on these APIs for the first time have no intuition for what a token costs them in practice. Here's the breakdown.
What's a token, roughly?
A token is a chunk of text — on average, about ¾ of a word in English. "The quick brown fox" is roughly 5 tokens. A 500-word blog post is roughly 650–700 tokens. A typical chat message might be 20–100 tokens; a long document could be tens of thousands.
Why input and output are priced differently
Every model charges more for output tokens than input tokens — often 4 to 8 times more. Generating text is computationally heavier than reading it. This matters a lot in practice: a chatbot that reads long documents but gives short answers is cheap to run; one that writes long reports from short prompts is expensive.
Current pricing for major models
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| GPT-5.2 | $1.75 | $14 |
| GPT-4.1 Mini | $0.40 | $1.60 |
| Claude Opus 4.6 | $5.00 | $25 |
| Claude Sonnet 4.6 | $3.00 | $15 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| Gemini 3 Pro | $2.00 | $10 |
| Gemini 3 Flash | $0.10 | $0.40 |
| DeepSeek V3.2 | $0.27 | $1.10 |
A worked example
Say you're building a customer-support assistant on Claude Sonnet 4.6. Each conversation reads roughly 2,000 tokens of context (chat history + knowledge base snippets) and writes a 300-token reply.
- Input cost: 2,000 tokens × $3/1M = $0.006
- Output cost: 300 tokens × $15/1M = $0.0045
- Cost per conversation: ~$0.0105
At 10,000 conversations a month, that's roughly $105/month — a number that's easy to miss when you're only looking at the per-token price and not multiplying it out against real volume.
Batch API: the discount most people miss
Nearly every provider now offers a "batch" mode — you submit requests asynchronously and get results within 24 hours, in exchange for roughly a 50% discount on both input and output tokens. For anything that doesn't need a live response (nightly data processing, bulk content generation, embeddings re-indexing), batch mode effectively halves your bill for free.
Context caching: the other discount
If you're sending the same large system prompt or document repeatedly, most providers offer cached input pricing — often 50–90% cheaper than standard input tokens for content the model has "seen" recently. This is one of the highest-leverage optimizations for RAG systems and coding assistants that resend the same context on every call.
Output tokens dominate the bill for most chat-style products. If your costs are higher than expected, check whether your model is writing longer responses than it needs to — that's usually the biggest lever, more than switching models.
Picking a model by cost per task
The cheapest model per token isn't always the cheapest per completed task — a weaker, cheaper model that needs three retries to get a correct answer can cost more than a stronger model that gets it right the first time. When comparing models on price, compare cost-per-successful-task, not just the sticker price per million tokens.