LLM Token Counter

Language-model APIs bill by the token — chunks of a few characters produced by a byte-pair-encoding tokenizer, roughly four characters or three-quarters of an English word each. This page runs the o200k_base tokenizer, the encoding OpenAI's current GPT models use, entirely inside your tab: the ~1 MB token table downloads once when you start typing and the text you paste is processed locally, never sent anywhere. For OpenAI models the count you see is therefore exact — the same number the API would meter.

The other providers publish no browser tokenizer, so their columns are estimates derived from the o200k count with calibrated multipliers, and the table labels them as such. The honest error bar is ±20–30%. Anthropic is the clearest case: its official position is that its last public tokenizer is no longer accurate for modern Claude models, and its pricing documentation notes that current Claude models tokenize the same text into roughly 30% more tokens than their predecessors — both facts this calculator's multipliers encode. For billing-grade counts on non-OpenAI models, use the provider's count-tokens endpoint or read the usage field the API returns with every response.

Take a 1,000-token prompt (o200k count) with a 500-token reply, run 100 times a day. On GPT-5 at $1.25 per million input tokens and $10 per million output, one request costs 1,000 × $1.25/1M + 500 × $10/1M = $0.00625, and the month lands at about $18.75. On GPT-5 nano the same traffic costs about $0.75 a month; on a frontier model at $10 input / $50 output — with the ~1.5× tokenizer adjustment taking the prompt to ~1,500 tokens — the request is roughly $0.04 and the month roughly $120.

That spread — two orders of magnitude for the same workload — is the point of the comparison table. Model choice dominates every other cost decision, and the calculator makes the comparison concrete for your actual prompt rather than a hypothetical one. The per-month column simply multiplies by your requests-per-day figure across thirty days; batch and caching discounts (next section) can cut it further.

Every major provider now discounts repeated input. OpenAI bills cached input at one tenth of the base rate automatically when a prompt prefix repeats. Anthropic's caching is explicit and asymmetric: writing a cache entry costs 1.25× the input rate (or 2× for hour-long entries) and reading it back costs 0.1× — so a cached system prompt pays for itself from roughly the first reuse, and everything after is a 90% discount on that prefix. Gemini adds a wrinkle: cache reads are cheap, but cached content also accrues a per-hour storage charge, so short-lived caches on high-traffic prompts are the winning shape there.

Batch APIs are the other lever: both OpenAI and Anthropic price asynchronous batch traffic at half the standard rate, input and output alike. A nightly classification job that tolerates minutes of latency should almost never run at real-time prices. The calculator's table shows standard rates — the cached-input column tells you what the discount applies to, and halving a row's per-month figure previews the batch price. DeepSeek adds time-of-day pricing (off-peak hours cost half its standard rate), which the table's footnote carries.

Token counts here cover text. Images, audio and PDFs are metered by separate per-item or per-resolution formulas that vary by provider, and chat requests carry a few tokens of per-message overhead beyond the raw text — small, but visible on short prompts. Reasoning models add another wrinkle: their internal thinking tokens bill as output on most providers, so a terse-looking answer can meter far more output than it displays. Treat the output-tokens box as an estimate you control, and check real usage from the API's own accounting once traffic flows.

Prices move. Every rate in the table was read from the provider's official pricing page on the date stamped above it, with the source links one click away; providers shift prices or add models every couple of months, and we re-verify against those pages on each update. Where a change is already scheduled — a promotional rate with a published end date, a provider switching billing models — the table's notes say so. If the stamp looks old, the linked pages are the authority.