LLM API Cost Calculator — Compare GPT, Claude, Gemini Pricing

What actually drives an LLM API bill

Output tokens. Output is priced 4 to 6 times higher than input at OpenAI, Anthropic and Google, and reasoning tokens are output. A workload that writes long answers or thinks hard is dominated by the output column, so a model with cheap input and expensive output can lose to one with the opposite profile.

Context you resend. Retrieval results, documents and chat history are input on every request. Ten thousand tokens of context on a million requests is ten billion input tokens, which at $2 per million is $20,000. Prompt caching cuts that by 90% or more when the context is identical across requests.

Tokenizer differences. Prices are per token, but tokens are not the same size across providers. Anthropic's current tokenizer produces about 30% more tokens than its previous one for the same text. Turn on Scale for each model's tokenizer to compare like for like.

Volume discounts that are not on the price list. Committed-use contracts, provisioned throughput and enterprise agreements can change the picture at large volumes. This calculator uses public pay-as-you-go prices.

Long-context price tiers, by provider

Several providers raise the price of the whole request once the prompt passes a threshold. The calculator applies these automatically and marks affected rows with Long-context rate.

• OpenAI: prompts over 272K input tokens on GPT-5.4, GPT-5.5, GPT-5.6 and GPT-6 models are billed at the long-context rate: double for input and 1.5x for output.

• Google: Gemini 3.1 Pro and 2.5 Pro charge more for prompts over 200K tokens ($4 instead of $2 input on 3.1 Pro). The Flash models and Gemini 4 Argon have a single rate.

• Anthropic: Claude 4.6 and later bill the full 1M-token context at the standard rate, with one exception: Claude Haiku 5.5 charges $0.50 instead of $0.10 per million input tokens once a prompt passes 100,000 tokens.

• xAI: every Grok model doubles its price when a prompt reaches 200K tokens, and the higher rate applies to every token in that request.

Reading the subscription vs API comparison

The subscription mode prices your chat usage as if you made the same calls through the API: messages per day, tokens per message and reply, and how long your conversations run. Turn k of a conversation resends the k-1 earlier messages and replies, so the average prompt grows with conversation length; the tool shows that average.

With Serve the conversation history from the prompt cache turned on, the part of each prompt that was already sent is billed as a cache hit and the new part as a cache write, which is how a well-built chat client on the API behaves. Turning it off shows the worst case.

The break-even column is the number of messages per day at which the API costs exactly the plan price. Below it, pay-as-you-go is cheaper. Remember what the comparison leaves out: plans include web search, file handling, image generation and a polished app, and they cap usage; the API has no app but no cap beyond rate limits.