| Model | Best for | Context | Per 1M tokens |
|---|
Flashdecuffed-qwen3.8-27bQwen3.8 · 27B | Fast everyday answers, writing and code. | 64K contextabout 49,000 words | about 450 questions per $1$0.50 in · $2.00 out per 1M · cached input 90% off |
|---|
Flagshipdecuffed-deepseek-v4.1-flashDeepSeek-V4.1-Flash · 552B / 16B active | The largest model in the catalog, and the fastest. | 256K contextabout 197,000 words | about 690 questions per $1$0.33 in · $1.32 out per 1M |
|---|
A typical question is 2,000 tokens in and 600 out. Thinking bills as output.
PLANSStart free. Pay when you need more.
FREETry it free
$0About 380 questions a month: 1 million tokens on any live model.
DEVELOPERA monthly allowance
$20 / monthEvery live model except Flagship at list prices, within two usage limits.
- $2.50 of use per five hours
- $11.50 of use per week
- Renews monthly until you cancel
PREPAIDPay as you go
$10 minimum top-upOne balance. Pay for the tokens you use at each model’s rate.
- Purchased credits don’t expire
- No automatic top-ups
- Credits can’t be transferred
Running production traffic? For teams.
Email and no-name accounts work too; they pay as they go or take the Developer plan.
AT LIST PRICEA month of real work, per model.
| Workload | Flash | Flagship |
|---|
| Chat | $2.73 | $1.98 |
|---|
| Coding agent | $4.80 | $7.92 |
|---|
| Long writing | $4.25 | $2.81 |
|---|
Before taxes and payment fees.
Flash’s price comes from its measured speed on one H100. See the numbers.
Mainstream APIs run $20–200 a month and their models refuse. decuff is pay-per-token, prepaid, and the models never refuse. The walls are enumerated.
How billing works
Free tokens cover any live model for accounts signed in with a phone, within daily limits. Then your Developer allowance, promotional credits and balance, in that order; a plan uses your balance only if you allow it. Past free tokens, only your balance pays for Flagship.
Cached input
Repeated prompt prefixes bill 90% off, except on Flagship. Each request reserves its maximum output, then charges what it used.
What “abliterated” means
Weights edited to remove learned refusals. No model answers every prompt or keeps every skill of its original.
What’s live right now: GET /v1/catalog on the API.