Beam 1.0 API pricing and cost calculator

How this is calculated

For a request with 100,000 input and 5,000 output tokens, uncached input costs $0.025 and output costs $0.005, so one call is $0.03. With half the input read from cache the call costs $0.02, and 100 calls a month come to $2. A reasoning model spends output tokens on its thinking before it answers, so output is the number to measure on your own tasks. In this example the call splits into $0.015 for input and $0.005 for output, and long thinking would move that split. Beam is aimed at long horizon agent work, so picture a session of 30 turns that each send 80,000 input tokens, mostly cached, and write 4,000 output tokens. That session costs $0.288 at the listed rates, and $0.72 if nothing is cached. Open weights add a second way to pay. A billion uncached input tokens at the hosted rate costs $250 and a billion output tokens costs $1,000. Those two figures are the yardstick for a hardware decision. Memory follows the 501B total parameters even though only 23B are active for each token, because the router can pick any expert. The VRAM calculator estimates that budget at 4 bit, and your own throughput measurements at your own context length turn it into a cost per token.

Verdict

Start with the hosted price and measure a real workload. Record input, cached input and output tokens per finished task, since a reasoning model's output can vary a lot between tasks. Then decide whether self hosting is worth the extra work. It only pays off when your own GPUs stay busy, so compare the cost of the hardware and the time to run it with the hosted figures above, using your measured throughput rather than a guess. Beam takes text only, so images and PDFs need converting to text first, and that step has its own cost. If a cheaper host appears, test it on your own tasks first, since the quantization it runs can change results. Read the Apache 2.0 license and the model card before you ship anything to production.