Pay for what you actually ran.
Every plan is a pot of credits. Each model has a published rate, output costs more than input, and the price of an answer appears on the answer. No per-seat surcharge for the expensive models, and no bill you only understand at the end of the month.
Invited
Enough to decide whether this is your tool.
Freewhile the invite lasts
300 creditsa month
- Haiku 4.5 and both DeepSeek V4 models
- Full reasoning ladder on models that have one
- Branching, prompt library, export
- Read-only share links
Pro
Most chosenThe whole lineup, metered to the token.
$20per month
2,000 creditsa month
- Every model — Claude, OpenAI, DeepSeek
- Fast mode on Opus 5
- Compare three models side by side
- Unused credits roll over one month
Max
For people whose day is spent in here.
$80per month
9,000 creditsa month
- Everything in Pro
- 12.5% more credits per dollar than Pro
- Compare four models at once
- Highest rate limits and concurrency
- Usage export and per-thread reporting
Run dry mid-month? Top up 1,000 credits for $10, any number of times. Top-up credits do not expire while your plan is active.
The rate card.
Credits per 1,000 tokens. Cheap models are genuinely cheap here — a long DeepSeek thread costs a fraction of a short Opus one — so picking the right model for the job is worth doing, and the app makes it a single click.
| Model | Reasoning | Input / 1K | Output / 1K |
|---|---|---|---|
| Claude Opus 5 | 5 levels | 1.00fast 2.00 | 5.00fast 10.00 |
| Claude Haiku 4.5 | None | 0.30 | 1.50 |
| GPT-5.6 Sol | 6 levels | 0.80 | 4.00 |
| GPT-5.6 Terra | 4 levels | 0.38 | 2.00 |
| DeepSeek V4 Pro | 2 levels | 0.16 | 0.57 |
| DeepSeek V4 Flash | 2 levels | 0.08 | 0.33 |
Caching does most of the work on a long thread: repeated context is charged at a fraction of the input rate, and the app shows your hit rate per thread. Rates change when providers change theirs; we publish before it takes effect.
Questions worth asking first.
- How are credits charged?
- Per token, at the rate published for the model you used — input and output priced separately, because they are not the same thing. The cost of an answer lands on the message itself, so nothing is reconciled at the end of the month.
- Why per token instead of per message?
- A message to Haiku and a message to Opus 5 at maximum effort differ by more than a hundred times in what they take to produce. Counting messages would mean charging everyone for the most expensive thing anyone does.
- Does reasoning effort cost more?
- Yes. Thinking tokens are output tokens, so the top of the ladder can spend several times what the bottom does on the same prompt. The rail shows you what you are asking for before you send it.
- What does fast mode do to the rate?
- Fast mode runs Opus 5 at up to 2.5x output speed and bills at double its normal credit rate. It is off by default, and the composer marks it clearly while it is on.
- What happens when I run out?
- Sending is refused before the request goes out, with the remaining balance named — never mid-answer. Top up at any point in the month; nothing already written is affected.
- Do unused credits expire?
- Pro and Max roll unused credits over for one month. Top-up credits do not expire while your plan is active.