LLM Model Pricing Update

Update src/data/llm_models.csv with the latest pricing information from each provider's official pricing page.

CSV Format

provider,model,input_price,output_price,max_input_tokens,max_output_tokens,pricing_url

• input_price / output_price: USD per 1M tokens
• max_input_tokens / max_output_tokens: integer token counts (no commas, no K/M suffix). Use - when not applicable (e.g., image-generation models) or when the spec cannot be confirmed from official docs.
• pricing_url: The provider's official pricing page URL
• Use - for output_price when it does not apply (e.g., image generation models)

Steps

1. Read the current CSV

Read src/data/llm_models.csv to understand the current list of providers and models.

2. Fetch pricing pages from each provider

Use WebFetch to retrieve the latest pricing information from the following URLs in parallel:

• OpenAI: https://platform.openai.com/docs/pricing

  • If the primary URL returns 403 (Cloudflare challenge), fall back to https://developers.openai.com/api/docs/pricing — it serves the same content and is fetchable.
  • Extract API prices (per 1M tokens) from the "Language models" section
  • Use standard API prices, not Batch API prices
  • Exclude models explicitly marked as deprecated/legacy
  • Include audio, realtime, search, computer-use, and image models
  • For multi-modality models (realtime, image), use the text input/output price for the CSV (audio/image pricing is tracked separately)

• Anthropic: https://www.anthropic.com/pricing

  • This URL now 301-redirects to https://claude.com/pricing, which does not include the API pricing table. Use the docs URL directly: https://platform.claude.com/docs/en/about-claude/pricing
  • Extract from the "Model pricing" table (Base Input Tokens column = input, Output Tokens column = output)
  • If prices are shown in MTok notation, they are already in per-1M-token units

• Google: https://ai.google.dev/pricing

  • Extract Gemini model prices (per 1M tokens)
  • Use Pay-as-you-go prices, not the free tier
  • Use the price for prompts up to 128K tokens when tiered pricing exists
  • For Live/multi-modal preview models with both text and audio rates, use the text rate

• DeepSeek: https://api-docs.deepseek.com/quick_start/pricing (preferred; https://platform.deepseek.com/pricing requires login)

  • Extract API prices (per 1M tokens)
  • Use standard prices, not cache hit prices
  • If a promotional discount is currently active, record the regular (post-discount) list price, not the promo price
  • Skip models the page marks as "will be deprecated" — list only the current/replacement model names

3. Fetch model specs (context window / max output tokens)

Pricing pages do not include token limits for OpenAI, Anthropic, or Google — fetch from each provider's model documentation. DeepSeek's pricing page already contains them.

• OpenAI: per-model pages under https://developers.openai.com/api/docs/models/<model> (e.g. .../models/gpt-5.5, .../models/gpt-realtime-2). Each page has "Context Window" and "Max Output Tokens". Image-generation models (gpt-image-*) do not publish these — use -.
• Anthropic: the comparison tables on https://docs.claude.com/en/docs/about-claude/models/overview (which redirects to https://platform.claude.com/docs/en/about-claude/models/overview). Read the "Context window" and "Max output" rows. Both the "Latest models comparison" and "Legacy models" tables are needed to cover all non-deprecated models.
• Google: per-model pages under https://ai.google.dev/gemini-api/docs/models/<model-slug> (e.g. .../models/gemini-3.5-flash, .../models/gemini-2.5-pro). Each page has "Input token limit" and "Output token limit". Note that slug conventions are inconsistent — some Preview variants and Live/Native-Audio models do not have individual pages; use - if not found.
• DeepSeek: already on the pricing page — "Max context length" → max_input_tokens, "MAXIMUM" output → max_output_tokens.

When a Preview variant lacks its own page, assume the same limits as the non-Preview variant of the same model family (and note this in the change report).

4. Update the CSV

Based on the fetched information, perform the following:

• Add new models: Add models that exist on the official page but are missing from the CSV. Group models by provider.
• Update prices: Update input_price / output_price for models whose prices have changed.
• Update token limits: Update max_input_tokens / max_output_tokens when the spec changes.
• Remove deprecated or deleted models: Remove rows for models that have been discontinued from the official page.
• Maintain the existing provider order (OpenAI → Anthropic → Google → DeepSeek).
• Order models within each provider according to the official page listing order.

5. Report changes

After the update is complete, report the following to the user:

• List of newly added models (include the token limits used and whether they were inferred from a sibling model)
• List of models with price changes (old price → new price)
• List of models with token-limit changes (old → new)
• List of removed models
• If no changes were made, report that as well

Notes

• If WebFetch fails or the page structure has changed, report the issue for that provider and prompt the user for manual verification.
• Skip and report models with unclear pricing (e.g., per-request billing only, not per-token).
• Ensure the CSV ends with a single trailing newline.