Menú
You maintain an internal **per-token LLM pricing data file** —
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Review code changes and report issues by severity with actionable fixes.
Sharpen a fuzzy intention into one measurable objective string that drives the rest of the work.
Convert a Prompt Flow PRS pipeline submission to run a Microsoft Agent Framework workflow.
Build a Model Context Protocol (MCP) server that lets an LLM call into external tools and resources.
Summarize PDF documents into concise bullet-point digests.
Bump a dependency version across a pnpm workspace and update lockfile.
| name | LLM Model Pricing File Update |
| description | You maintain an internal **per-token LLM pricing data file** — |
| category | finance |
| tags | ["ai","api","backend","cli","deployment"] |
| source | {"url":"https://github.com/langfuse/langfuse/tree/8624bbf82b06381c461c84e75bdb5fc6626dbb1a/.agents/skills/add-model-price","fetched_at":"2026-06-12","commit":"8624bbf82b06381c461c84e75bdb5fc6626dbb1a","license":"MIT","original_path":".agents/skills/add-model-price/SKILL.md"} |
| license | MIT |
| author | Langfuse (downstream pack: badhope) |
| version | 0.1.0 |
| needs_review | false |
| slug | llm-pricing-file-update |
| created | 2026-06-12 |
| updated | 2026-06-19 |
| inputs | [{"name":"pricing_file","type":"string","required":true,"description":"Absolute path to the JSON price file"},{"name":"action","type":"string","required":true,"description":"Action type - add (new model) or update (existing)"},{"name":"model_spec","type":"object","required":true,"description":"Model data including id, prefixes, provider, tokenizer, tiers, source URL"},{"name":"types_file","type":"string","required":false,"description":"Path to TS module listing selectable models"}] |
| output | {"format":"markdown","description":"Generated content based on the user request"} |
You maintain an internal per-token LLM pricing data file — a JSON your product reads to convert raw token usage into USD for cost dashboards, billing, and budget alerts. New model releases, price drops, and tokenizer changes mean the file is never done.
The work splits into three concerns:
gpt-4.1-turbo arrives
under many names: gpt-4.1-turbo, openai/gpt-4.1-turbo,
azure/gpt-4.1-turbo-2025-04, gpt-4.1-turbo-2025-04-15.
Your matchPattern regex needs to catch the canonical id
and the per-provider variants without catching lookalikes
(gpt-4.1, gpt-4.1-turbo-mini).Don't use this skill for the API call to the model itself (that's normal client code), or for invoice reconciliation (that's a separate process against the bill from your provider). And don't use it for cost allocation across teams — that's tagging, not pricing.
| Field | Required | Notes |
|---|---|---|
pricing_file | yes | Absolute path to the JSON price file. |
action | yes | add for a new model, update for changes. |
model_spec | yes | The model data: id, prefixes, provider, tokenizer, tiers, source URL. |
types_file | no | TS module listing selectable models (if applicable). |
Plain-text confirmation. Typical return:
Diff applied to worker/src/constants/default-model-prices.json:
+ gpt-4.1-turbo
matchPattern: (?i)^(openai/|azure/)?(gpt-4\.1-turbo)(-[0-9]{4}-[0-9]{2}-[0-9]{2})?$
tokenizer: o200k_base
tiers: [{ input: 2.50, output: 10.00, cache_read: 1.25 }]
updatedAt: 2026-06-10
Diff applied to packages/shared/src/server/llm/types.ts:
+ "gpt-4.1-turbo"
Regex self-check:
accept: gpt-4.1-turbo, openai/gpt-4.1-turbo, azure/gpt-4.1-turbo-2025-04
reject: gpt-4.1, gpt-4.1-turbo-mini, gpt-4.1-turbo-32k
all 3 / 3 accept, 0 / 3 false-positive
Schema validation: pass
Downstream consumers that may need a refresh:
- playground model selector
- eval cost rollup job
- monthly budget alert thresholds
You are updating a per-token LLM pricing data file. Follow the
order. Do not skip the regex self-check — a regex that matches
nothing is silent death.
1. Decide: add or update?
`add` — the model id does not appear in the file at all
`update` — the id exists, but prices, tiers, regex, or
tokenizer change
2. For `add`: gather the model spec.
- Canonical id (what the provider calls the model in their
official pricing page).
- Provider prefixes: list the strings your platform will
see in the wild. At minimum:
- bare: `gpt-4.1-turbo`
- with prefix: `openai/gpt-4.1-turbo`
- per-provider: `azure/gpt-4.1-turbo-2025-04-15`
- per-dated: `gpt-4.1-turbo-2025-04-15`
- Provider: openai | anthropic | bedrock | vertex | azure
| gemini. Drives which pricing conventions apply
(cache read vs cache write, batch discount, etc.).
- Tokenizer: the actual tokenizer name the model uses,
not the model name. `o200k_base`, `claude`, `gemini`,
`p50k_base`. Mismatches here silently under- or
over-count tokens.
- Source URL: the official pricing page. REQUIRED.
Do not invent prices.
3. Write the entry.
Shape (one entry per model):
{
"id": "gpt-4.1-turbo",
"matchPattern": "(?i)^(openai/|azure/)?(gpt-4\\.1-turbo)(-[0-9]{4}-[0-9]{2}-[0-9]{2})?$",
"tokenizer": "o200k_base",
"tiers": [
{
"name": "default",
"input_cost_per_1m_tokens": 2.50,
"output_cost_per_1m_tokens": 10.00,
"cache_read_cost_per_1m_tokens": 1.25
}
],
"updatedAt": "2026-06-10"
}
Rules:
- `matchPattern` MUST anchor with `^` and `$` so
`gpt-4.1` does NOT match `gpt-4.1-turbo`.
- Use `(?i)` for case insensitivity unless the
provider is strict-case.
- Always include the dated variant as optional
(`(-[0-9]{4}-[0-9]{2}-[0-9]{2})?$`).
- At least one tier, named `default`.
- `updatedAt` is today's ISO-8601 date.
4. For `update`: only change what changed.
Do NOT rewrite the entry. Update prices, regex, tokenizer,
or tier names in place. Always refresh `updatedAt`.
5. Update the types file (if applicable).
If your product has a TypeScript union / enum of allowed
model ids, append the new id. Do not alphabetize for
the sake of alphabetizing — preserve the order of
existing entries; new entries go at the end.
6. Run the regex self-check.
Use the supplied tester. Pass at least:
- 3 accept samples (canonical, prefixed, dated)
- 3 reject samples (predecessor, suffix variant,
unrelated model in the same family)
If any accept fails → regex is wrong, fix it.
If any reject passes → regex is greedy, tighten it.
7. Validate the JSON file.
Run the file's validator (a node script in your repo
that re-parses the JSON and checks tier invariants:
non-negative prices, `input < output` is NOT enforced —
some providers charge more for input, e.g. Claude
Opus with prompt caching, but the validator should
check `>= 0`).
8. List downstream consumers.
Ask: "If I add `gpt-4.1-turbo` to the pricing file, who
else needs to know?"
- Playground model selector (UI)
- Eval cost rollup job (uses prices to convert usage
to USD)
- Budget alert thresholds (manager-configured per
team / per project)
- Self-hosted customers' dashboards (need to pull
the new prices via upgrade)
Do not auto-update these. Just list them so the user
can decide.
9. Commit the diff in one PR.
Pricing file change + types file change (if any) +
validation script run + a short CHANGELOG entry.
Do not split the change across two PRs — the pricing
and the type must land together or the build is
broken.
tiers to mark it as deprecated, set prices to zero,
and let the dashboard filter it out..*. That is a footgun. A regex
without anchors matches everything. The validator
will (or should) reject unanchored patterns.claude tokenizer on a Gemini model will
under-count, and o200k_base on an older p50k_base
model will over-count. The pricing file is only as
good as the tokenizer alignment.Input:
pricing_file: worker/src/constants/default-model-prices.json
action: add
model_spec:
id: gpt-4.1-turbo
provider_prefixes: [gpt-4.1-turbo, openai/gpt-4.1-turbo, gpt-4.1-turbo-2025-04-15]
provider: openai
tokenizer: o200k_base
tiers:
- name: default
input_cost_per_1m_tokens: 2.50
output_cost_per_1m_tokens: 10.00
cache_read_cost_per_1m_tokens: 1.25
source_url: https://openai.com/api/pricing/
types_file: packages/shared/src/server/llm/types.ts
Output:
Diff applied to worker/src/constants/default-model-prices.json:
+ gpt-4.1-turbo
matchPattern: (?i)^(openai/)?(gpt-4\.1-turbo)(-[0-9]{4}-[0-9]{2}-[0-9]{2})?$
tokenizer: o200k_base
tiers: [{ input: 2.50, output: 10.00, cache_read: 1.25 }]
updatedAt: 2026-06-10
Diff applied to packages/shared/src/server/llm/types.ts:
+ "gpt-4.1-turbo" appended to ModelId union
Regex self-check:
accept (3/3 pass):
gpt-4.1-turbo → match
openai/gpt-4.1-turbo → match
gpt-4.1-turbo-2025-04-15 → match
reject (3/3 pass):
gpt-4.1 → no match
gpt-4.1-turbo-mini → no match
gpt-4o → no match
Schema validation: pass
- JSON parses
- all prices >= 0
- all matchPatterns anchored
- updatedAt is a valid ISO-8601 date
Downstream consumers that may need a refresh:
- frontend playground model selector
- eval cost rollup (nightly job)
- budget alert thresholds (manual config)
- customer-facing changelog (1-line entry: "+gpt-4.1-turbo")
These are the bugs that bite every new user. Check them before shipping:
Regex without anchors: matchPattern without ^ and $ matches unintended models.
Invented prices without source URL: Using made-up numbers because official pricing isn't published yet.
Tokenizer mismatch: Using wrong tokenizer for a model silently miscounts tokens.
Updating file without updating types: TypeScript types drift from pricing file.
Silent fallback to zero cost: Model not matching any pattern gets $0 cost.