llm-price-update

name	llm-price-update
description	Refresh LLM provider pricing (OpenAI, Anthropic, Google Gemini) in configs/base.yml for the llm-proxy app. Use when the user asks to "update llm prices", "refresh model pricing", "sync llm-proxy pricing", "check llm pricing", "llm price update", or otherwise wants the llm-proxy pricing config audited against current vendor pricing.

llm-price-update

Audit and update model pricing in configs/base.yml for the llm-proxy app against the current published prices from OpenAI, Anthropic, and Google. Produce both the pricing diff and a deprecation list.

Hard rules

1. Always use the live web. Do NOT rely on model knowledge for prices or model lineups. Prices change frequently; the agent's training data is stale.
2. Always dispatch parallel subagents (one per provider) to gather pricing. Then verify the most surprising claims with direct WebSearch calls before writing the config. Subagents can hallucinate; independent verification is required for any new model family, price drop >20%, or price increase.
3. Never add aliases for models that are actually different versions. Aliases are only for the same underlying model (e.g., claude-opus-4-1 ↔ claude-opus-4-1-20250805, gpt-4o ↔ gpt-4o-2024-11-20). Never alias gpt-5 to gpt-5.1, claude-opus-4-0 to claude-opus-4-1, gemini-2.5-pro to gemini-3-pro, etc. Different version numbers = different models = separate entries.
4. Always encode tiered pricing when the vendor publishes it. This is not optional. If the vendor lists different input/output rates based on prompt length (e.g., Gemini Pro ≤200k vs >200k, Gemini 1.5 ≤128k vs >128k, Anthropic Sonnet 4.x ≤200k vs >200k), the model MUST be configured as a PricingTier list — never collapsed to flat {input, output}. Flat pricing silently under-bills long-prompt traffic.
5. Always consult each vendor's official deprecations page. Do not rely on the pricing page alone — models often linger on the pricing page after sunset, and new sunsets are announced on the deprecations page first. Cross-reference every model in configs/base.yml against the URLs in "Vendor URLs" below, and surface any retired/sunset model in the deprecations output.
6. Always produce a deprecations list alongside the pricing diff so it can be consumed by downstream callers (e.g., client UI, docs). See "Deprecation output" below.

Workflow

Copy this checklist at the start of a run:

Task progress:
- [ ] 1. Read configs/base.yml to capture current pricing and models
- [ ] 2. Dispatch 3 parallel subagents (OpenAI, Anthropic, Google), each reading BOTH the pricing page AND the deprecations page
- [ ] 3. Verify surprising findings with direct WebSearch
- [ ] 4. Diff vendor data against configs/base.yml (pricing + deprecations)
- [ ] 5. Build pricing changeset + deprecations list (including shutdown models to remove)
- [ ] 6. Present plan (or apply, depending on mode)
- [ ] 7. Run `go run ./cmd/config-validator/` after editing
- [ ] 8. Run `go test ./internal/config/...` after editing

Step 1. Read the current config

Only configs/base.yml holds pricing. dev.yml, staging.yml, and production.yml inherit pricing from base and do not override it. Confirm with a grep before assuming:

rg -n "pricing" configs/

The pricing struct is defined in internal/config/config.go as ModelPricing with three shapes:

• {input, output} for flat pricing — ONLY when the vendor publishes a single rate for the model
• [{threshold, input, output}, ...] for prompt-length tiers — REQUIRED whenever the vendor publishes length-based tiers (Gemini Pro ≤200k/>200k, Gemini 1.5 ≤128k/>128k, Anthropic Sonnet 4.x ≤200k/>200k, etc.)
• overrides: {alias: {input, output}} for per-snapshot price differences (e.g., gpt-4o-2024-05-13 priced higher than current gpt-4o)

When building the diff, explicitly check each new model's vendor pricing page for "long context", ">200k", ">128k", "prompts above X tokens", or a second pricing row. If any exists, encode it as tiers.

Step 2. Dispatch parallel subagents

Launch all three in one tool-call batch using subagent_type: generalPurpose. Each gets the current list of models in its provider from configs/base.yml, and is told to return per-million-token standard-tier prices (non-cached, non-batch) plus any new models released since the config was last updated.

Prompt skeleton (adapt per provider):

Read BOTH <vendor pricing URL> AND <vendor deprecations URL> and return two tables:

Table 1 — current pricing. For every model in the list below, the current per-million-token input and output prices in USD for the standard (non-cached, non-batch) tier. Include any tiered pricing (e.g., >200k input tokens) as an explicit second row. Cross-reference against a second source (OpenRouter, artificialanalysis, vendor docs) if possible.

Models in current config: <list from base.yml>

Also list ANY NEW <vendor> models released since <assumed cutoff> that we should be aware of. For each new model: official API model name (NOT aliases), input/output $/1M, tier threshold, release date.

Table 2 — deprecations. Every model on the vendor's official deprecations page, including models NOT currently in our config. Columns: model, status (deprecated / shutdown / legacy), sunset_date (ISO or null), replaced_by, source URL, notes.

Format both tables as compact markdown tables. If a model appears on the pricing page but NOT the deprecations page, treat it as active. If it appears on the deprecations page but NOT the pricing page, treat it as sunset and include it only in Table 2.

Vendor URLs — the subagent MUST read both the pricing page and the deprecations page for its provider:

Provider	Pricing	Deprecations	Secondary
OpenAI	https://platform.openai.com/docs/pricing	https://developers.openai.com/api/docs/deprecations	https://openai.com/api/pricing/
Anthropic	https://docs.anthropic.com/en/docs/about-claude/pricing	https://platform.claude.com/docs/en/about-claude/model-deprecations	https://www.anthropic.com/pricing
Google	https://ai.google.dev/gemini-api/docs/pricing	https://ai.google.dev/gemini-api/docs/deprecations	https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions

The deprecations page is authoritative for sunset dates and replacements. The pricing page is authoritative for current rates. Both must be read.

Step 3. Verify surprising findings

For any subagent claim that looks suspicious — new major version (e.g., "GPT-6"), price change >20%, model name you don't recognize — run a direct WebSearch with 2-3 queries to confirm. If verification fails, drop the claim and note it in the plan.

Step 4. Build the diff

For each provider section of configs/base.yml, categorize every model:

Category	Action
Unchanged	Skip
Price changed	Update input/output; if an alias has a different price, move it into overrides
New model	Add entry with canonical ID + same-model aliases only (see rule 3)
Deprecated (future sunset)	Keep in config, add to deprecations list with sunset_date
Shutdown (past sunset, vendor 404s on it)	Remove from configs/base.yml, add to deprecations list. Google retired gemini-1.5-* and text-embedding-004/embedding-001 this way.
Listed as deprecated on vendor page but pricing page still shows it	Keep in config, add to deprecations list with sunset_date if known

Reuse the limits block of a similar existing model rather than inventing new rate-limit numbers — e.g., a new flagship uses the same limits as the previous flagship.

Step 5. Deprecation output

In addition to the pricing diff, emit a machine-readable deprecations list the user can copy elsewhere (client UI warnings, release notes, etc.). Format:

deprecations:
  openai:
    - model: o1-preview
      replaced_by: o1
      sunset_date: null  # ISO date if vendor announced one, null if only delisted
      status: deprecated  # deprecated | shutdown | legacy
      source: https://developers.openai.com/api/docs/deprecations
      notes: Removed from platform pricing page
  anthropic:
    - model: claude-3-5-sonnet
      replaced_by: claude-sonnet-4-5
      sunset_date: null
      status: deprecated
      source: https://platform.claude.com/docs/en/about-claude/model-deprecations
      notes: Dropped from marketing page; still listed on docs pricing
  gemini:
    - model: text-embedding-004
      replaced_by: gemini-embedding-001
      sunset_date: "2026-01-14"
      status: shutdown
      source: https://ai.google.dev/gemini-api/docs/deprecations
      notes: Shutdown Jan 14, 2026; API now 404s
    - model: gemini-2.0-flash
      replaced_by: gemini-2.5-flash
      sunset_date: "2026-06-01"
      status: deprecated
      source: https://ai.google.dev/gemini-api/docs/deprecations
      notes: Scheduled shutdown June 1, 2026

Include this block in the plan/PR body and, if the user asks, write it to a file (suggest docs/model-deprecations.yml or similar — do not create it unprompted).

Step 6. Present or apply

• In Plan mode: call CreatePlan with the diff + deprecations block. Surface open questions (e.g., tiered pricing, same-price aliases vs separate entries).
• In Agent mode: edit configs/base.yml directly, then go to step 7.

Step 7. Verify after editing

Always run both commands before committing:

# 1. Semantic config validation (pricing sanity, duplicate aliases, structural correctness)
go run ./cmd/config-validator/

# 2. Unit tests (config parsing, tiered pricing logic, alias resolution)
go test ./internal/config/...

Also grep for any hardcoded price assertions in tests that touch changed models:

rg -n "\"<changed-model-id>\"" internal/ configs/

Update test expectations if tests assert specific prices that changed. The validator and test suite must both be clean before the PR is considered ready.

Alias rules (non-negotiable)

The governing principle: same version number = aliasable, different version number = never aliasable.

An alias is acceptable when any of the following is true:

• It is a dated snapshot of the same model (e.g., claude-opus-4-1-20250805)
• It is a "latest" pointer for the same model (e.g., claude-3-5-sonnet-latest)
• It is a punctuation variant of the same model (e.g., claude-opus-4.1 ↔ claude-opus-4-1, gemini-3.1-pro ↔ gemini-3-1-pro)
• It is a preview / experimental SKU for the same model version (e.g., gemini-2.5-pro-preview under gemini-2.5-pro, gpt-4.1-2025-04-14-preview under gpt-4.1). Preview aliases are fine as long as the version number matches.

Examples of valid aliases:

• claude-opus-4-1 aliases claude-opus-4-1-20250805 (dated snapshot)
• gpt-4o aliases gpt-4o-2024-11-20 (latest snapshot pointer)
• gemini-3.1-pro aliases gemini-3.1-pro-preview and gemini-3-1-pro (preview SKU + punctuation variant, same version)
• claude-sonnet-4-5 aliases claude-sonnet-4.5 and claude-sonnet-4-5-20250929

Examples that are NEVER aliases (always separate entries), because the version number differs:

• gpt-5 and gpt-5.1 — different versions
• gpt-5 and gpt-5.2 — different versions, even if priced identically
• claude-opus-4-0 and claude-opus-4-1 — different versions
• claude-opus-4-5 and claude-opus-4-6 and claude-opus-4-7 — each is its own entry
• gemini-2.5-pro and gemini-3-pro-preview — different version families
• gemini-3-pro-preview and gemini-3.1-pro-preview — 3 and 3.1 are different versions
• o1 and o1-pro — different products
• Any model and its -mini/-nano/-lite/-pro sibling

When in doubt, create a separate entry. Over-aliasing silently routes traffic to the wrong model's pricing and rate limits, and masks deprecations.

Snippet templates

Flat pricing (most common):

"<model-id>":
  enabled: true
  aliases: ["<dated-snapshot>", "<vendor-documented-same-model-alias>"]
  limits:
    # copy from sibling model of same tier
    tokens_per_minute: 80_000
    requests_per_minute: 1_000
    tokens_per_day: 1_920_000
    requests_per_day: 1_440_000
    burst_tokens: 8_000
    burst_requests: 100
  pricing:
    input: <in>
    output: <out>

Tiered pricing (Gemini Pro, Anthropic >200k):

"<model-id>":
  enabled: true
  aliases: []
  limits: { ... }
  pricing:
    - threshold: 200_000
      input: <low-tier-in>
      output: <low-tier-out>
    - threshold: 0  # fallback for prompts above threshold
      input: <high-tier-in>
      output: <high-tier-out>

Per-snapshot override (snapshot priced differently from the alias base):

"gpt-4o":
  pricing:
    input: 2.50
    output: 10.00
    overrides:
      "gpt-4o-2024-05-13":
        input: 5.00
        output: 15.00

Output format

At the end of the run, produce two artifacts:

1. Pricing changeset — a list of edits grouped by provider, with old vs new prices. Flag each entry as fix-stale, price-change, new-model, remove-shutdown, or no-op.
2. Deprecations list — the YAML block above, sourced from each vendor's official deprecations page. Always include it, even if empty (emit deprecations: {openai: [], anthropic: [], gemini: []}). Entries with status: shutdown SHOULD correspond to remove-shutdown entries in the changeset.