Frequently asked questions
How the bill is actually computed
Two numbers are metered separately and added: input tokens (everything you send (system prompt, conversation history, retrieved documents), tool results) and output tokens (what the model generates). Per-token, output is the expensive one, across the 226 models we track, the median output-to-input ratio is 4.0×, and the widest is 12.0×.
And yet on agent-shaped workloads the input side dominates the invoice. That is not a contradiction, it is arithmetic: an agent re-sends its entire context on every turn, so a turn that produces 800 output tokens may carry 60,000 input tokens. At a 4× unit price ratio, the input still accounts for roughly 95% of that turn's cost. Optimising output length on this workload is optimising the wrong end.
| Model | Input / 1M | Output / 1M | Output ÷ input | One agent turn | Input share of that turn |
|---|---|---|---|---|---|
claude-sonnet-5 | $1.00 | $5.00 | 5.0× | $0.064 | 94% |
gpt-5.6-luna | $0.100 | $0.600 | 6.0× | $0.006 | 93% |
deepseek-v4-flash | $0.150 | $0.600 | 4.0× | $0.009 | 95% |
The practical reading: unit price decides, but token shape decides harder. A model that costs twice as much per token can still be cheaper end to end if it needs half as many turns to finish the task, and a prompt that carries 40k tokens of irrelevant history costs the same as one that carries 40k tokens of useful context. Before switching models, measure which side of the sum your workload actually lives on.
When this is the wrong move
This is not the right answer for every workload. Specifically:
| Situation | Why it breaks | Do this instead |
|---|---|---|
| Regulated or highly sensitive data | Any third-party route adds a party that sees the payload. No rate advantage offsets that. | Contract directly with the vendor under your own data-processing terms. |
| You need a signed availability commitment | A compatibility route is an extra hop with no SLA of its own. | Keep the critical path on a contract you can enforce. |
| You have not measured your own token split | Every comparison on this site depends on the input/output ratio; the wrong ratio inverts the answer. | Meter tokens per task first, then compare rates. |
| Your volume is too small to matter | Below a few dollars a month, the effort of switching exceeds the saving. | Optimise later; spend the time on the workload shape instead. |
Is this actually the same model, or a cheaper substitute?
The same model. You call the identical model ID through an OpenAI-compatible endpoint. If you ask for claude-opus-5 you get Claude Opus 5, with the same context window. Verify it yourself by running a prompt you know the answer to before you trust it with real work.
How much cheaper is it?
It depends on the model. The Claude family and most of the GPT-5.6 family currently run at 50% of documented official list price. GPT-5.6 Sol is 37.5% off input and 25% off output, because OpenAI's own $4/$20 rate is promotional. Some models are at or near list. We publish the current per-model rate rather than a single headline number.
Do I have to prepay?
No. The minimum top-up is $1 and the balance does not expire. This is the main difference from prepaid credit bundles, which typically start at $250 and go up to $10,000.
How do I pay?
Card via Stripe, plus additional methods depending on region. Balances are denominated in USD.
Which models are actually discounted?
Verified against vendor docs: the Claude family (Opus 5, Sonnet 5, Fable 5, Haiku 4.5) and GPT-5.6 Terra / Luna / GPT-6 Astra run at 50% of documented list. GPT-5.6 Sol is 37.5% / 25% off. For DeepSeek, GLM, Kimi, Qwen and MiniMax we publish the gateway rate but claim no discount, because those vendors bill by peak/off-peak windows and cache tiers, one percentage would be misleading.
Does it work with Claude Code, Cursor, Cline and Aider?
Yes. All of them let you override the API base URL. See the setup guides for Claude Code, Cursor and the SDKs.
What happens if the gateway goes down?
Your requests fail. Keep your direct provider key configured as a fallback and put the base URL in an environment variable so you can switch back without redeploying.
Is this allowed by Anthropic and OpenAI?
These routes generally sit in a grey area relative to vendor terms of service. They are widely used, but they are not an official reseller channel. If you need contractual guarantees, buy direct.
Can I use this for production?
Many teams do. Start with non-critical traffic, verify output quality on your own tasks, and keep a fallback configured. Regulated or sensitive data should not go through any third-party gateway.
How do I know these prices are current?
Every price on this site is checked against the vendor's own pricing page and the gateway's live rate config. Each page that lists rates carries the date those rates were verified; pages that list no rates carry no date. Prices change without notice, so confirm before you commit a workload.