LLM API Cost & Token Budget Estimator

How token billing actually works

Providers bill per token, and a token is roughly four characters of English, or about 750 words per 1,000 tokens. The exact count depends on the content and the tokeniser, so treat word-based estimates as ±20% rather than precise.

The important structural fact is that input and output are priced differently, and output is usually several times more expensive. On the models in this table the output price runs between four and five times the input price. Input is processed in parallel during the prefill phase; output is generated one token at a time, auto-regressively, which is far more compute-intensive per token.

That single ratio drives most cost engineering. Anything you can move from output to input, or eliminate entirely, saves disproportionately — which is why shorter answers and better prompting usually beat switching models.

The two numbers people underestimate

Prompt length. A typical bare question is 50 to 200 tokens. The same question inside a real application — system prompt, retrieved documents, tool definitions and conversation history — commonly runs 1,500 to 8,000 tokens before the user's actual words. Applications do not get billed for the user's question; they get billed for everything they send alongside it.

Thinking tokens. Reasoning models generate intermediate tokens, and those are billed at the OUTPUT rate even though users never see them. Google states this explicitly for Gemini ('output price, including thinking tokens'). A reasoning model can therefore cost several times a plain chat model on identical visible output, and the only way to see it is to log reported token usage rather than estimating from response length.

About these prices, and how to keep them honest

Every price in the table was read from the provider's own pricing page on 2 October 2026, and each row links to the page it came from. The verification date is shown in the results, because a price table without a date is a guess wearing a suit.

Prices change quietly. Introductory rates expire — several Gemini Flash tiers step up materially on 1 January 2027, and OpenAI's promotional rate on GPT-5.6 Sol is guaranteed only to 21 November 2026. DeepSeek publishes peak and off-peak rates where off-peak is exactly half. Anthropic splits long-context and fast-mode pricing. So the calculator lets you overtype any input or output price with your own contract rate; the defaults are a starting point, not an authority.

One row is marked as derived rather than read: Claude Sonnet 5's input and output cells were blank in Anthropic's published table, so the figures shown are reconstructed from its three cache-rate multipliers, which independently imply a $2.00 input rate. It is flagged in the results so you can check it rather than trust it.

Discounts that change the answer completely

• Cached input. A cache hit typically costs a tenth of the standard input rate. Anthropic charges 1.25x for a 5-minute cache write and 2x for a 1-hour write, then 0.1x per read; OpenAI publishes a cached-input column at roughly 0.1x. A stable system prompt behind a cache is the single largest saving in most applications.
• Batch processing. Non-urgent work submitted in batch is discounted — Google publishes a straight 50% reduction, and OpenAI prices batch separately. Anything that does not need a live answer belongs in a batch.
• Service tiers. Priority and fast tiers cost more, sometimes double. Reserve them for genuinely latency-sensitive paths.
• Model routing. Sending every request to a frontier model is the most common way to overspend. Most production traffic is classification, extraction or short generation that a cheaper tier handles at a fraction of the cost.

Costs this estimator does not include

It models token spend for the request shape you describe. Real bills also carry:

• Tool and search fees, billed per call on top of tokens. OpenAI charges $10 per 1,000 web search calls plus the retrieved content as input tokens; Google gives 5,000 grounding requests a month then charges $14 per 1,000. An agent that searches on every request can spend more on search than on inference.
• Retries and failures. Rate-limited or malformed requests that are retried still cost tokens on the successful attempt.
• Image, audio and video tokens, which are priced on entirely different scales and are not comparable per token with text.
• Storage and compute outside the model — vector databases, container minutes for code execution, file storage for uploaded assets.
• Fine-tuning, which providers increasingly price separately from inference.

How to read the comparison table

The table ranks every model by total monthly cost for the traffic you described, and the spread is usually wide — often two orders of magnitude between the cheapest and the most capable option. That spread is the useful information.

Start from the required capability and work down, not from the cheapest model up. Decide what quality the task actually needs, pick the cheapest model that clears that bar, then use routing so only the requests that need more get more. A single model choice for an entire application is almost always either overpaying or underperforming on part of the traffic.

Before committing, put a hard monthly spend cap and alert thresholds in the provider dashboard. Runaway cost in an agent loop is a real failure mode, not a hypothetical one, and the caps are free.