What a month actually costs
Per-token prices are hard to reason about until you turn them into a monthly bill. The table below does that for three usage sizes, with cached input priced separately — because for a coding agent, cached input is most of the input.
The three profiles
Light
5M input + 1M output per month, half the input served from cache. Roughly an individual using an agent for part of the week.
Steady
30M input + 6M output per month, 70% of the input served from cache. Roughly a developer using an agent daily.
Heavy
150M input + 30M output per month, 80% of the input served from cache. A small team, or one long-running agent.
| Model | Light | Steady | Heavy |
|---|---|---|---|
claude-sonnet-5 | $14.72 | $78.09 | $364.80 |
claude-opus-5 | $36.81 | $195.23 | $912.00 |
claude-haiku-4.5 | $7.36 | $39.05 | $182.40 |
gpt-5.4 | $20.78 | $111.86 | $527.25 |
deepseek-v4-pro | $1.96 | $9.55 | $42.15 |
deepseek-v4-flash | $0.63 | $3.07 | $13.57 |
kimi-k2.7-code | $6.51 | $34.71 | $162.73 |
qwen3-coder-plus | $14.40 | $77.76 | $367.20 |
glm-5.2 | $6.16 | $31.31 | $142.42 |
minimax-m2.7 | $1.54 | $8.00 | $36.94 |
Monthly cost in USD at current catalog prices. Cached input is billed at the cache-read price; the rest of the input at the input price.
Plug in your own numbers
The arithmetic is not hidden. Take your monthly token counts, split the input by how much of it is repeated context, and apply the per-million prices from the catalog.
monthly cost
monthly cost =
(input_M x (1 - cached_share) x input_price)
+ (input_M x cached_share x cache_read_price)
+ (output_M x output_price)
# input_M / output_M = millions of tokens per month
# prices = USD per 1M tokens, from the catalogWhy the cache share matters more than the headline price
Cached input bills at 10% of the input price, and cache writes carry no surcharge — the first pass bills at the normal input price and nothing extra. On a workload where most of the input is repeated context, that moves the bill far more than a few percent off the sticker price does.
That is why the three profiles above assume rising cache shares: the heavier the agent use, the more of the context is repeated, and the wider the gap between the naive per-token estimate and what you are actually billed.
These figures are estimates from the published catalog prices and the stated assumptions — not a quote. The console shows the live price applied to each request and your actual usage.
FAQ
Where do these prices come from?
The same catalog the Models page publishes, rebuilt from the production price book whenever it changes. No page carries hand-typed prices.
Is there a subscription or seat fee on top?
No. You pay per token in USD from prepaid credits. There is no seat, no monthly minimum, and no subscription window.
What if my cache share is lower than assumed?
Use the formula above with your own split. If you are not caching at all, set the cached share to zero — the whole input bills at the input price.
Create a key and the console will show your real usage against these prices from the first call.