AI usage

Trylle-hosted AI is metered by model and token direction. Every model has input, output, and cache-read prices. A cache-write price applies only when that provider separately bills cache creation.

What consumes Trylle AI credit

• Platform automations that run a hosted model.
• Product features routed through the Trylle AI gateway.
• Bot or automation runs whose configuration selects a hosted provider through Trylle.

Direct provider calls from try ask, try explain, or try draft use your locally configured provider credentials. They are not charged to the Trylle wallet.

ZDR models

Models carrying the ZDR badge use a dedicated zero-data-retention route. ZDR usage follows the same funding order as every other hosted model: included AI credit, eligible promotional credit, then wallet balance.

Cost formula

For each dimension:

cost = token_count / 1,000,000 × model_rate

The final charge is the sum of input, output, applicable cache-write, and cache-read costs. Cache reads are usually cheaper than uncached input, while cache writes can cost more when a provider bills them.

For DeepSeek, Moonshot AI, Z.ai, and Grok, a cache hit means a repeated prompt prefix was read from the provider's cache. Trylle records those tokens as cache reads and does not apply a cache-write charge. OpenAI and Anthropic models retain their provider-specific cache-write dimensions where applicable.

Priority and fast modes use their published rates only when the provider reports that mode as effective. If the response reports a standard/default tier, standard rates apply.

Example

Suppose a gpt-5.6-terra run uses:

• 80,000 uncached input tokens
• 10,000 output tokens
• 20,000 cache-read tokens

input:     80,000 / 1,000,000 × $1.00 = $0.08
output:    10,000 / 1,000,000 × $6.00 = $0.06
cache read:20,000 / 1,000,000 × $0.10 = $0.002
total:                                      $0.142

Control spend

• Use smaller models for classification, routing, and high-volume routine work.
• Keep reusable prompt prefixes stable so cache reads can be reused.
• Limit repository context to files and diffs the task actually needs.
• Review failed automation runs: model work already performed can still be billable.
• Configure wallet auto-refill only after setting a threshold and refill amount appropriate for the account.

See Pricing for the current model catalog.