PromptGate computes the dollar cost of every request at write time using a per-(provider, model) input + output rate from the model pricing catalogue. The catalogue auto-refreshes daily from the community-maintained LiteLLM pricing JSON and accepts manual overrides that survive every subsequent refresh.

Where the cost shows up

• Every row in gateway_logs carries a cost_usd column populated at write time.
• The Cost Dashboard (/projects/{uuid}/cost) sums these values for the 24-hour / month-to-date / 30-day rollups.
• The Budget Enforcer uses month-to-date cost_usd to enforce monthly_budget_usd caps.
• Per-request cost is shown in the Playground response footer.

How a row is picked

Resolve order, given (provider_key, model) from a request:

1. Manual exact match — (provider, model) + source = manual.
2. Auto exact match — (provider, model) + source = auto.
3. Longest-prefix fallback — a row whose model is a prefix of the request’s model, longest first. E.g. gpt-4o-mini covers gpt-4o-mini-2024-07-18.
4. Per-endpoint fallback — the endpoint’s own estimated_cost_per_1k_tokens_usd (blended total-tokens rate). Preserves the pre-catalogue behaviour.
5. Null — no cost logged for the row; Cost Dashboard reads it as zero.

Admin: /admin/pricing

The page is split into three sections:

Section	What it does
KPI row	Total rows, last-refresh status + age, count of unmapped models, schedule cadence.
Unmapped models	(provider, model) pairs seen in gateway_logs over the last 30 days that have no pricing row. “Map prices” pre-fills the modal with the pair locked, so you only set rates.
Pricing catalogue	All rows, paginated. Edit any auto row to flip it to manual (immune to future refreshes). Add a brand-new manual entry.

The Refresh now button at the top right runs php artisan pricing:refresh on demand. Every run is captured in the Recent refresh runs disclosure with seen / added / updated / skipped counts.

Refresh internals

The command php artisan pricing:refresh fetches the JSON, then for each row:

1. Reads litellm_provider and maps it to one of our provider keys: openai, anthropic, google (covers gemini, vertex_ai-*), mistral, groq, cohere, together, ollama. Anything else → skipped.
2. Strips the provider prefix off the LiteLLM key (groq/llama3-70b-8192 → llama3-70b-8192).
3. Converts input_cost_per_token and output_cost_per_token (per token) to per-million for human-readable storage.
4. Upserts an auto row. manual rows are never overwritten — operator overrides win permanently.

Schedule lives in routes/console.php:

Schedule::command('pricing:refresh')->dailyAt('04:30')->withoutOverlapping();

Manual entries

Click + Add manual entry to register a custom row. Use it for:

• Enterprise contract rates that differ from public pricing.
• Models LiteLLM hasn’t catalogued yet (new releases, fine-tunes, private deployments).
• Mapping unmapped models seen in your logs.

Once saved, the row carries source = manual and is shown with a bright manual chip. Editing an existing auto row also flips it to manual.

Data model

model_pricing
  id, provider_key, model,
  input_per_1m_usd, output_per_1m_usd,
  source ('auto' | 'manual'),
  notes,
  created_at, updated_at
  UNIQUE (provider_key, model)

model_pricing_refreshes
  id, source_url, status,
  models_seen, models_added, models_updated, models_skipped,
  error_message,
  started_at, finished_at

API

The catalogue is operator-only for now — no public API surface. The Management API (/api/v1/control/*) does not yet expose pricing endpoints. If you need to bulk-load rates from IaC, run php artisan pricing:refresh from your provisioning step or seed the table via a custom seeder.