Cost tracking

Cost is calculated after each response from the driver’s rates and stored in ai_runs.cost (USD). Without rates cost is null, but token counts are always saved.

Rates

Per driver in config/ai-tasks.php, per 1M tokens:

'anthropic' => [
    'model' => 'claude-sonnet-5',
    'price' => [
        'in' => 3.00,
        'out' => 15.00,
        'cache_write' => 3.75,
        'cache_read' => 0.30,
    ],
],

A missing cache_read/cache_write falls back to in, so cached tokens are never costed as free. To cost them as free, set the rate to 0 explicitly.

All price keys, including per_char for TTS and per_minute for transcription — Configuration.

Per-model rates

price is one set of rates per driver, while the model comes from .env — switching to a pricier model silently keeps costing the old rates. List the exceptions under prices, keyed by model name; anything not listed falls back to price:

'deepseek' => [
    'model' => env('DEEPSEEK_MODEL', 'deepseek-flash'),
    'prices' => [
        'deepseek-reasoner' => ['in' => 0.55, 'out' => 2.19, 'cache_read' => 0.11],
    ],
    'price' => ['in' => 0.15, 'out' => 0.60, 'cache_read' => 0.003],
],

Keys match both the full model name and the part after /, so a gateway-prefixed anthropic/claude-sonnet-5 also matches a claude-sonnet-5 entry.

Rates snapshot

The rates used are stored with the run in ai_runs.cost_rates:

{"model": "deepseek-flash", "source": "driver", "in": 0.15, "out": 0.6, "cache_read": 0.003}

source is model:<name> when the rate came from prices, driver when from price. cost is computed from config at run time, so without the snapshot a row written before a price change or a model switch can’t be explained afterwards. It also makes drift detectable: recompute a period from tokens at today’s rates and compare with stored cost.

The column comes from a migration added in 3.27.0 — after upgrading run vendor:publish --tag=ai-migrations and migrate. php artisan about shows whether the schema is up to date; until then runs are stored without the snapshot.

Tokens

Column	Contains
tokens_in	Input tokens billed at full price only — cached ones are never included
tokens_out	Output tokens, reasoning included (billed at the output rate)
cache_read_tokens	Input read from the provider’s prompt cache
cache_write_tokens	Input written to the prompt cache

The meaning is the same for every driver.

Prompt caching

Providers that cache automatically (OpenAI, Gemini implicit cache, DeepSeek, Mistral) report cached tokens, and they are costed at the cache_read/cache_write rates.

Note

Anthropic’s explicit cache breakpoints (cache_control) are not exposed by the package yet. The 'cache' => true payload option shown in older READMEs has no effect.

Querying spend

use Fomvasss\AiTasks\Models\AiRun;

AiRun::where('tenant_id', $tenantId)
    ->whereBetween('created_at', [now()->startOfMonth(), now()->endOfMonth()])
    ->whereNotNull('cost')
    ->sum('cost');

Filter by cost IS NOT NULL, not status = 'ok': a run rejected by the post-call budget check was billed by the provider but has status error. This is how budgets count spend too.

Known approximations

• DeepSeek doubles its rates during peak hours; cost has no time-of-day tiering, so it’s understated for calls in those windows.
• TTS (per_char) is an estimate from the input length — the provider returns no usage.