LLM pricing calculator

Input and output are different products

Every provider bills the tokens you send and the tokens the model writes at separate rates, and output is always the expensive one. Input is read in a single parallel pass over the whole prompt. Output has to be generated one token at a time, each one a full forward pass through the model, and that serial work cannot be batched the way reading can. The ratio lands between 3x and 6x almost everywhere: Claude Sonnet 5 is $2 in and $10 out, GPT-5.6 Sol is $5 in and $30 out, Gemini 3.1 Pro is $2 in and $12 out.

The practical consequence is that two workloads with the same token count can cost very differently. A summarizer that reads 50,000 tokens and writes 500 is cheap. A code generator that reads 5,000 and writes 4,000 is not, even though it moved a tenth as many tokens. Estimate the two halves separately or the answer will be wrong by an order of magnitude.

Cached input is the biggest lever you control

When you send the same prefix twice, the provider can reuse the work it already did. Anthropic prices a cache read at 0.1x the input rate on every Claude model, and OpenAI does the same on the GPT-5.6 family. On a coding agent that carries the same system prompt, tool definitions, and repository context into every turn, that prefix is often 80 to 95 percent of the prompt, so caching takes most of your input bill to a tenth.

Writing to the cache costs more than a plain input token, which is the part people miss. Anthropic charges 1.25x input for a five-minute cache and 2x for a one-hour cache. A five-minute cache therefore pays for itself after a single read, and a one-hour cache after two. The failure mode is a cold loop: restarting a session per task pays the write premium every time and never collects the read discount, which is exactly the leak Continuum Scan looks for.

Long-context surcharges are a cliff, not a slope

Several providers charge a premium once a single request crosses a prompt-size threshold, and the premium applies to the whole request, not just the tokens past the line. Gemini 3.1 Pro doubles from $2 to $4 input above 200K tokens. Every Grok 4.x model uses its standard tier at 200K prompt tokens or fewer and doubles above 200K. GPT-5.6 Sol goes from $5 / $30 to $10 / $45 on a long-context request. Anthropic is the exception: Claude 4.6 and later carry the full 1M window at standard pricing, so a 900K-token request bills at the same per-token rate as a 9K one.

This turns context management into a pricing decision. A prompt that sits at 195K tokens and one that sits at 205K are 2x apart on most providers, so a retrieval step that trims context can be worth more than switching models.

Batch and off-peak discounts are free money for async work

If a job does not need an answer in the next few seconds, it should not be paying interactive rates. Anthropic's Batch API is 50 percent off both input and output, and it stacks with caching. DeepSeek does the same thing on a clock instead of an endpoint: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, and every other hour is exactly half price. Nightly evals, backfills, and classification runs are the obvious candidates.

Reasoning tokens bill as output, and tokenizers differ

Thinking or reasoning tokens are charged at the output rate even though you never see most of them. An effort dial left at maximum for routine work is billed in full, which is why the same prompt can cost three times more on one setting than another with an identical visible answer.

Finally, a token is not a fixed amount of text and not the same across families. Claude Opus 4.7 introduced a new tokenizer, shared by Opus 4.8, Opus 5 and Fable 5, that produces roughly 30 percent more tokens for the same input than earlier Claude models. Comparing two models on price per million tokens alone therefore understates the newer Claude models slightly, because the same document is more tokens. Count your real prompt against the model you plan to use: our token counter gives a fast estimate, and each provider publishes an exact endpoint.