Token Cost Explained: What You're Actually Paying For

A token is about 0.75 words, or roughly 4 characters of English text, and you pay per million of them: from $0.05 per 1M input tokens on DeepSeek V4 Flash to $5 on GPT-5.5 and Claude Opus 5. A 1,000-word document is about 1,333 tokens, so sending it to GPT-5.4 ($2.50 per 1M input) costs about $0.003.

Tokenisation is how models process text: words are converted into numerical tokens before processing. Examples:

• "The quick brown fox" = about 4 tokens
• "Hello, world!" = 3 tokens
• A typical short email = 100–200 tokens
• A 1,000-word blog post = about 1,333 tokens

Token counts vary by model. GPT models use the tiktoken library; Claude, Gemini and DeepSeek use their own tokenisers. Non-English text, code and symbols usually tokenise less efficiently, so budget 20–30% more for those.

Input tokens are what you send. Output tokens are what the model writes back. Output is almost always dearer, often by several times.

OpenAI GPT Latest example:

• Input: $2.00 per 1M tokens
• Output: $10.00 per 1M tokens
• A 1,000-token response costs $0.01 in output — a fraction of a cent

That asymmetry is why “be concise” is a cost control, not a style note. Capping output length is usually the fastest per-call saving available to you.

Work in millions of tokens per month, not per call. Per-call numbers are too small to reason about and hide the total.

Example calculation:

• You send 500 requests/day, each with 500 input tokens and 200 output tokens
• Monthly input: 500 × 500 × 30 = 7.5M tokens × $2.00/1M = $15.00
• Monthly output: 500 × 200 × 30 = 3M tokens × $10.00/1M = $30.00
• Total: $45.00/month on OpenAI GPT Latest

Same workload on OpenAI GPT Mini Latest:

• Monthly input: 7.5M × $0.75/1M = $5.63
• Monthly output: 3M × $4.50/1M = $13.50
• Total: $19.13/month — 57% cheaper

And on a high-throughput budget row such as DeepSeek V4 Flash ($0.04/$0.08 per 1M), the same month lands near $0.54. The model you route to matters far more than the prompt tricks you apply to it.

Every model has a context window — the maximum number of tokens it can process at once. Most current flagships list 1M tokens: GPT-5.4, GPT-5.5, Claude Sonnet 5, Claude Opus 5 and Gemini 3.1 Pro. DeepSeek V4 Flash and Llama 4 Scout list 1.3M; Claude Haiku 4.5 lists 200K; GPT-5.4 mini and nano list 400K.

A big window is a cost risk, not a cost saving. Every token in the window is billed as input on every request, so an agent that keeps 300K tokens of files in context pays 300,000 × $2.50 / 1M = $0.75 per step on GPT-5.4 — $15 for a 20-step run before any output is counted.

Optimisation: send only the last N messages of history, summarise older context instead of including raw text, and let agents retrieve files on demand rather than preloading a whole repository.

Anthropic, OpenAI and Google all offer prompt caching — re-using the processing of an identical prompt prefix. Cached input reads cost roughly 10% of the list input rate on Anthropic and OpenAI and roughly 25% on Gemini; Anthropic charges about 125% of list to write the cache the first time.

Worked example: a 5,000-token system prompt on Claude Sonnet 5 ($2 per 1M input) costs $0.01 per request uncached. Cached, each later read is roughly $0.001. At 10,000 requests a month that is about $10 instead of $100.

Two more discounts worth knowing:

• Batch API: Anthropic and Google list batch rows at 50% of the standard rate (Claude Sonnet 5 batch: $1 / $5 per 1M; Gemini 3.8 Flash batch: $0.375 / $1.875). Use it for anything that can wait hours — backfills, evaluations, nightly classification.
• Tool-call loops: an agent re-sends its context on every step, so caching the stable prefix (system prompt, tool definitions, loaded files) is the single largest saving for agent workloads.

Most production applications should enable caching by default and route non-interactive work to batch.